whosaid — who said what, by pipeline read-only

0:00Quinn SlackJust building software to do X is only valuable if you can get people to, to pay you. And I don't, I think that that's gonna be a lot harder given that people can just make the software themselves. So developers need to think of themselves as doing X, not building software to do X. Yeah. And I think that's gonna be a rude shift cuz if anything, software developers on average have shifted to be more general, more horizontal in what they're building.

0:22Thorsten BallThe richest companies on earth are pouring all of their cash into this. And we're not even seeing the results of this yet. Like, this is still being built. If you now think, oh, I have to be stingy with my monthly subscription and I cannot run out of tokens, I think in a year or two, you will find that the tokens are everywhere. It's like electricity. It's like you're not measuring the watts coming out of your wall. And I think it's the same thing with tokens. There will be so many tokens everywhere and so fast. that people won't recognize computing anymore. And we're back. Another episode of Raising an Agent. This time we're back home. One of us in San Francisco, in Mill Valley, the other one in this really small town in Germany near Frankfurt Airport. And yeah, we're excited to talk again, right? Like Munich sparked a lot of ideas. Munich got us get the stuff flowing in our heads, the jelly flowing.

1:28Quinn SlackYeah, we've been shipping a lot. And it's been a few weeks since Munich. And we had said then that the way that we build had changed more in the prior, like, 4 or 6 weeks than it had changed all of last year. And I think we're still feeling that. And we're starting to think about looking even further ahead and what is the world going to be like and How can we take what we're building, what we're seeing, and get more people building in the AMP way? And also then thinking about what is that going to do to software and what that means for software, right?

2:03Thorsten BallYeah.

2:04Quinn SlackSo, Thorsten, predictions. What do you think is going to be different 6, 9 months from now?

2:10Thorsten BallI think all of the trajectories will continue. Like, I don't think there's gonna be this rug pull that most people say there might be or there will be. So I think the price of intelligence will drop. And people are saying, well, the models are more expensive. That's not— well, the newest ones, right? But like, the price of the old models has dropped considerably. And I think in, say, half a year, the price of GPT-5, 6 levels of intelligence will be dirt cheap compared to right now. And, you know, Fable, incredibly expensive. I think intelligence at that level will be cheap. I also think if you look at the trajectory of tokens per second, right, OpenAI is rolling out their super fast mode. I don't have access. I wish I had access. It's, what is it, 750 tokens per second, which is mind-blowing.

3:06Quinn SlackIf we had access, we would not be able to say.

3:08Thorsten BallSo, if I had access, I wouldn't be able to say. be here talking, to be honest. I would just send tokens. That's mind-blowing. So I think this is— I'm not sure if this is correct, actually, but I think it's close to like the build-out we're seeing with AI, the data centers, the GPUs. It's one of the biggest infrastructure build-outs in the history of humanity. You know, like this— the richest companies on Earth are pouring all of their cash into this. And we're only just— we're not even seeing the results of this yet. Like, this is still being built. So, if you now think, oh, I have to be stingy with my monthly subscription and I cannot run out of tokens and I count the tokens and, you know, 50K tokens is something that I measure, I think in a year or 2 or maybe 3, you will find that the tokens are everywhere. It's like electricity. You're not measuring the watts coming out of your wall, and you won't measure. You know, you measure them in fucking megawatts and kilowatts, and you don't care about like 50 watts, right? And I think it's the same thing with tokens. There will be so many tokens everywhere and so fast that people won't recognize computing anymore. I think based on these projections, a lot of things that we think about how software development is will change. Well, do you disagree with any of these predictions or do you wanna add some?

4:40Quinn SlackI agree. Intelligence is getting cheaper and cheaper. Yeah. And anyone who thinks there's gonna be a rug pull, the quote unquote subsidies will end. There's no evidence that subscriptions, which is what people usually mean, are actually unprofitable. And it's hard to actually tease it apart in the same way that It's not like United Airlines is going to stop having economy seating because that's been a subsidy and they're going to pull the subsidy away. No. Once there's a big upfront expense, it makes sense to spread that in a price-discriminated way across the largest possible population of buyers.

5:15Thorsten BallYeah.

5:16Quinn SlackAnd also, totally agree that if you are in a role where you're having to be really stingy with the intelligence that you can use, then that's a signal in much the same way. If you're someone who's getting paid 1/20th of what a San Francisco engineer makes, then find a way to get to San Francisco or closer to San Francisco. It's a signal that what you're doing is not benefiting enough from tokens. You're not creating enough value. You're not in that good position. But if you're thinking that, then you're probably the kind of person who can create a ton of value if you get into a job or other setting where— the tokens can flow. And at the same time, there's a lot of noise going on right now. First, it feels like everybody is trying to build the same thing. Now, you know, Thurston, you and I have been building dev tools for a long time. And for the longest time, everyone said, oh, you can't make any money in it.

6:14Thorsten BallYeah.

6:15Quinn SlackAnd now people have completely changed to the point where it feels like everybody is building dev tools. And that also is probably not correct.

6:25Thorsten BallYeah.

6:26Quinn SlackAnd ultimately, software, it has to solve the problem. Just building software to do X is only valuable if you can get people to pay you. And I think that that's going to be a lot harder given that people can just make the software themselves. So, I think more and more, we've said this before, but developers need to think of themselves as doing X, not building software to do X.

6:50Thorsten BallYeah.

6:50Quinn SlackAnd I think that's gonna be a rude shift, 'cause if anything, software developers on average have shifted to be more general, more horizontal in what they're building.

6:58Thorsten BallI do have a little pet theory that I wanna offer here, and it's a pet theory. Say the inflection point was November 25th, right? Which, this is my other pet theory. It wasn't for us. I don't think November 25th was that big. I don't think there was a bump in the curve. I think it was a, Except for Gemini, remember? That was— well, let's not talk about it. But I think it was a smooth curve. But there's a lot of engineers who've kind of woken up to agents now. And now we've seen like in the last few months, people— and I've said this before on this podcast, but it's even more than this where people are now like they're hacking on their own agents and like the plugin and it's I thought this would go away, but it's still a thing that's going on where people build their own tooling around this stuff and blah, blah, blah, blah, blah. And my pet theory is engineers have this thing now where they go, it can't be this easy, right? Like, you know, like how we always try to fix our tools and build like some elaborate workflows and automate stuff. I mean, not me. I'm the guy who copy and pastes this. You know, I'm not doing this, but Like, if you send a prompt to Fable, I did this yesterday. I used Fable. I'm kind of on a 2-day Fable binge. I don't know. Maybe it's because I want to spend money after we can put ChatGPT subs into AMP. And now I'm like, well, you know, gotta use Fable. So I sent a prompt yesterday. Ian Landsman, shout out to Ian. He sent a tweet. He's like, I want to be able to set the default branch in AMP for orbs. So I sent a prompt. I tweeted the prompt. It was like a text message.

8:39Quinn SlackI sent...

8:40Thorsten BallI would send to somebody. And Fable went and did it. And it just added a feature. It didn't add complexity. It kind of wired that new field, the default branch, through our different services. I had it watch deployment. I had it make sure it rolled out. I tested it, and it works. I do think, and here's the pet theory, that engineers then go and look at this and go, oh, it can't be this easy, right? How about I add a bunch of stuff? I add some tools and some shortcuts and a review thing and this process and that process and blah, blah, blah. Because it doesn't appeal to the engineers of the last, say, 20 years that you just say something in plain English and it goes and does it. You know, like, I think there's like something like this going on where you want to kind of add like some value and that's how you build like these other engineering, this whole elaborate stuff around it.

9:32Quinn SlackBut yeah, most of the time— Well, it's like if you're a furniture builder, you would probably feel ashamed if someone comes over your house and it's all IKEA furniture. And when you sit down on that IKEA couch, you're gonna hear every little creak and it's gonna annoy the hell outta you.

9:45Thorsten BallI mean, we all know this story, right? That some guy posted, I don't know, 20 years ago on the internet. They were a 3D designer and basically they figured out that every time they had like something they made and they showed it to their boss and said like, does this look good? The boss would always say, just to— the boss would feel they have to add value. So they would ask like, change this, even though it was completely nonsensical. So the designer added a little duck to everything, and then the boss would always go, that looks great, but get rid of the duck. And then they would feel happy about it. They added something. So I do think there might be something going on here where engineers feel like it can't be as easy as just saying something to the model and build software, right? And so they have to add something. Yeah.

10:34Quinn SlackSo with an agent though, you wrote the How to Build an Agent that turned on a lot of people to how simple the idea of an agent loop with tools could be. And that was pretty cool. That opened people's eyes. That's a new primitive. What about all these people who say, oh, orbs, how is that different from just EC2 and you set up a VPN and you have some orchestration? You know, like what are they going to run into that they can't build themselves?

11:05Thorsten BallThere's multiple things going on. I think, well, first of all, we have to admit that building stuff is pretty easy now and getting started with stuff, right? And we had like the first version of what's now called Orbs in 2 weeks, a week or something. And then we ran into the long tail with that infrastructure provider we used and of issues and in a whole slog. And then I think I rewrote it on the new platform over the course of a weekend. And this was not me sitting here for 12 hours. This was me starting prompts and whatnot. And so admittedly, it is easy to build a lot of stuff. But when somebody says, oh, I built Orbs and it's easy to, and a VPN, or I have 15 tmux sessions running on my home lab and 18 different Claude codes in it, and— You can get far, but it's obviously not the same experience. The value of Orbs is that you have all of the things in the same package. You don't have to manage tmux sessions. The beauty of Orbs is you don't have to think about the machine. You just send your prompt, you get the results back, you preview it in a portal, you can share it with your colleagues and whatnot. You don't have to worry about how the sausage is made. And There's also people who are offended by this where they're like, I want to see how the sausage is made, you know? But I don't know. I don't know what to say. Like, it's the same thing maybe. Like, are you really concerned about how we spawn up these sandboxes or VMs? Like, do you— what do you want to see? You know, like you can specify hooks and scripts now. You can define your own bash script. We're pretty exposing a bunch of stuff, but I think it's underselling the whole experience. Like you can get like 80% of the way there, but What we found is the 20% makes a lot of difference. It's weird how hard it was to get this nailed down because it's not like super hard programming. Also, what does that mean anymore? I'm not going on this sidetrack, but you know, like if you'd asked me 5 years ago, super hard programming would be like an optimizing pass for an optimizing compiler, you know? I'm pretty sure that has kind of flipped now. Like—

13:19Quinn SlackYeah, it's like the design. And yeah, the design, all these pieces fit together.

13:24Thorsten BallYeah, into a whole. And like the product, what are you building? In what sequence? When are you building it? What are the trade-offs? You know, all of this stuff. So when we built this, we've— this was really hard. Like how— what are the primitives, for example? And what are the things that— the invariants of the product? One thing that I think we got right is You start an orb, you get a single URL, you get a fresh orb for new threads. That was controversial when I said this, right? Everybody was like, no, no, no, I want to spawn multiple agents in an orb, in the same thing. And now look at us. That would be a different product where you can spawn multiple agents and select a specific orb and whatnot. That would be a different thing. So, people can go far and wire their own stuff together, but wiring your own stuff together is not building the same product. And there's a lot of decisions that are not code level that go into making the full thing. And those seem to be valuable, right? If you look at the feedback we get where people say it just works or it's a great experience, or somebody tweeted this just an hour ago, like, your UI or your frontend is completely on a different level and whatnot, right? And obviously, we have great designers and engineers on the team who really sweat over the details here. And I don't know. I don't know. I don't know how you can replicate this with your homemade VPN EC2, the Linux thing.

14:59Quinn SlackWhat do you think? I think there's 2 ways to look at it. One is if you're just trying to have fun or just build something for yourself to use and you don't need to make it work at like team scale, you don't need to get other humans involved, you don't need to deal with the complexities that come with other people also wanting to make it work their way, then, you know, you could probably build something like an agent platform that's pretty fun.

15:25Thorsten BallYeah.

15:26Quinn SlackBut once you start making it so that an entire team can work on it, that all the different ways they use it work, that you can read other people's threads, you can have the, you know, Puck-like orchestrator agent work across everything in it, that all the concepts fit together, Yeah. That's hard. And the hard part for us has been no single person on our team, you know, our team is 20 people, all of AMP is 20 people. No single person on our team has every single piece of the system and the design in their head. And so much of it is figuring out how we can have all the different pieces that we all know and feel really strongly about work well together.

16:05Thorsten BallYeah.

16:06Quinn SlackAnd that is tough. We all use the product all the time. We have a small team that's highly trusted. So we've optimized AMP for being able to do that. But that's where if it's, you know, you don't really have that need if you're one person building it for yourself. And if you're one person building it for your whole company, you're going to run into all the same problems that we have. And then people are going to be saying, well, why can't I use AMP or, you know, all the other— now there's 87 different things that are in this category. And that is probably going to give you a headache. But the more important thing is like, what is— it's fun. Don't get me wrong. That's part of why we do it. But you're not actually creating value. It's not like any of these agents like Amp. It's not like we're charging you a ton for the joy of using our product. We're making money. We're profitable. But you're not going to save a lot of money. You're probably going to actually lose money on net because there's a lot of efficiency stuff we do. So, what is the business value you're doing? You can't be making software to do X. You most certainly cannot be making software to make software to do X. You gotta be doing X these days.

17:13Thorsten BallYeah. Yeah. But I mean, it points at a, at a real tension, right? The whole build versus buy thing, right? Where should we buy something and not build it ourselves? And 2 years ago, you know, besides operational costs and like guarantees and support and whatnot, you had the big thing of, I don't want to build this myself, or, I can't build it myself. But now, in the age of, I take a screenshot and I say, build this for me, that equation has changed. That scale has been tipped. And I think the interesting bit is that the models are now weighed on the scale, on the build versus buy scale, and the weight of the models changes constantly. You know, like you constantly have to reevaluate. And I mean, admittedly, I built a bug tracker into Amp last week, right? And it was one-shot and it was deeply integrated into our system. So that might've been even easier than spending the time to setting up Linear or something because now the value is in deeply integrating stuff. Which is why we're even talking here, right? Like this whole thing is incredibly fascinating to me that if you look at the history of software, And we've said this 1,000 times, software was hard and expensive to write, slow. Basically, you want the investment you make when writing software to pay off or have a return. So you generalize the software that you write so you can sell it or reuse it multiple times in different use cases, right? You don't build one-off software. That's not— doesn't pay or it hasn't paid. But—

18:55Quinn SlackThe beauty of a software business is you build it, And then for the next 10 years, your customers will stay on it and you can sell it and infinite copies can be made at zero marginal cost and you can make money on each of those. So you recoup your fixed expense.

19:09Thorsten BallYeah. Even if you're an agency and you build customer-specific things, most agencies have then developed some kind of framework or like some, I guess you can call it a stock or whatever, like something that they reuse across projects, right? Which means they can save time like a common admin interface or whatever. Whatever. Or think of like WordPress, right? Which is, I mean, it's been 10 years since I checked, but back then it was the most configurable piece of software. Like the admin interface had a lot of menus. And now it's like, well, you could take a screenshot or 4 of different websites and have like one-shot a thing, like your hair salon website or like your whatever restaurant website. And then every time you wanna make a change, you can ask the agent again to do it, right? So this has changed. And like what I find fascinating is that a lot of practices and methodologies of software, we still do them or they're kind of crumbling, but a lot of stuff doesn't hold anymore, right? Where we talk about—

20:14Quinn SlackYeah, and so before you even get onto that, I think you mentioned this, but I just wanna, put it to you again. Basically what we are saying is we see a lot of people say, oh, AMP or other agents, I could just replace that with EC2 and some shell scripts and get the same thing. And, and we understand that, but we don't think that that's the right thing for them to do. But at the same time, you're saying I was able to code the parts of Linear, an issue tracker that I need from scratch.

20:41Thorsten BallYeah.

20:41Quinn SlackI, you know, quote unquote replaced Linear with a few CRUD forms. So, which is correct? How could we say that in one case, you know, AMP is worth it, but in the other case, no, linear is not worth it? What makes it different?

20:55Thorsten BallThat's a good question. I think what's really different about— let's start on this end. I would say the linear or bug report case is so different because the workflow around how you do bugs is different in each company. And it's like bug tracking or issue reporting or knowledge management. It's different in every company. That's why Jira has 15,000 options, right? And why every Jira is different. It's not because people don't like the button, but it's because they work in a different way. So, I think like—

21:35Quinn SlackCould you say the way that people use agents and do software development is different in every company and therefore every company has a different—

21:40Thorsten BallI mean, I guess you could. I'm starting to— I'm thinking about it. But I would argue that it's like the thing I said at the start, right? Like, you have like, you need to specify what you want and the model does the thing. And there's not a lot of this. It's handed over to the model is, I guess, what I'm saying. There's not a lot of touchpoints anymore with the humans. Maybe that's the difference in that, you know, like previously, like your dev environment had like a lot of touchpoints where the humans were changing things. But now Man, like in Orbis, like they just whip up software and tear it down and I don't care. And I don't know, what do you think? Like, this is interesting. What do you think?

22:19Quinn SlackI think this is the big question that everyone has about software is what software is below the waterline where you don't want to go dive and mess with it? And what software is above the waterline where you want to, you know, go build? If you just look at people's behavior, I think you've seen a lot of people shift from wanting to spend their time in issue land and project land and typing that in. And now they want to spend their time in agent land. And you cannot argue with people's revealed preferences, their actual actions. And so, there's something about, well, why do we think AMP or generally the agent is different from the issue tracker? It's because we have found it to be incredibly productive to work there, far more than issues ever would. And so there's, there's more value to having that be really good and solid.

23:14Thorsten BallOh, I see what you're saying. So you, okay, I see what you're saying. So the value, the business value of the software has changed and you would say that building your own for low value things is, that's, that's, you know, we don't spend that much time in it ourselves.

23:32Quinn SlackSo, there's not that much complexity that we need to replicate. So, in the past, maybe we would've used 60% of the features of Linear. And Linear is an awesome, awesome product. You know, we've used it and is the best of its category.

23:49Thorsten BallYeah.

23:49Quinn SlackBut now we might only be using 10%. And I think there's a lot of people that are using some of Linear's agent features and they like that, but we would only be using 10% of it. So I get, I think there's also an element where Linear was built for a workflow that is undergoing a lot of change. And if you have, I think there's also something fundamental about software businesses that were started before 2025 and have not completely reinvented themselves and burned the bridges. I think there's also a lot of valid skepticism that, that those workflows will not be be there. I think there's also something about if like an agent is at the core, that then that just makes the software so much more pliable and moldable. Like in the way with AMP, if you wanna do something, most of the time you can just ask Puck to do it. If you want an integration, you can just ask, you know, Puck or AMP in a thread to go set up a webhook. And yeah, I don't know, but you know, this, this is what we're thinking about right now.

24:48Thorsten BallYeah. Another angle is the whole, I don't really like the term, but like the whole system of record type of thing, meaning like some software is more important than others. I think that's kind of what people want to say when they use the term, meaning with orbs, you're spending money on VMs. So as soon as you're spending money, like significant amounts of money, that changes this whole equation, right? And you have the tokens flowing through the system, you want to keep track of this. I do think like the whole process that touches the code is important. So, you know, like there's something systemy or recordishy about like the orbs, whereas with the tickets thing that I built, it's not, right? Like for us, that's not a cost. Like it's a CRUD app in our admin CRUD app. It's basically free. And that changes it for me. And like the whole orb stuff, that's our main application. That's the main thing that's That's load-bearing, to use the stolen term the agents stole from us. But I don't know, maybe that's this. But I do think like what you said is very interesting with you have the agent at the core and you can change it. And there's obviously now this tension of Linear was built in a time where you had to, for different customers, they have different workflows. So you had to have the option to Custom labels or like custom workflows or like custom boards and whatnot. And admittedly, Linear was, I think, three years ago they were pretty minimal, right? And they said like we only have this. And then people said like the follow. I don't know. That's not true. But like a couple years back, they started adding more features and it became more complex, complex. And obviously, if you if you're SaaS software, you have to be this complex because you have multiple customers and they give you different amounts of money and they have different. Usage and different workflows. So your software has to be generic for different things, right? But with agents now, if you build your own, it doesn't have to be generic at all. And that kind of changes the whole cost thing of the software again. Like it makes it less complex and whatnot. But obviously, like what I'm talking about here is like this triangle of, well, you know, you can buy the SaaS thing and it's complex in the backend and a lot of work to build because you have to generalize it, or you build your own and then it's kind of maybe cheap to build but it's customized to your workflow. But then on the other end of the triangle, there's this thing that, you know, with orbs and whatnot where it's load-bearing, it costs money, you do want to modify it, but yeah, I don't know. Like, I think everybody's trying to figure out where— what goes where on this triangle.

27:33Quinn SlackYeah. Yeah. And with AMP, we kill features that we no longer think are good. So there's not cruft in AMP that we're keeping around because some customer 9 months ago paid us for a year and they'd be pissed if we take it out. No. I mean, everyone knows with AMP that there's no cruft in it. And by the way, Linear is, of basically any software company, best positioned here. And they've done way better than pretty much any other company I know in staying relevant. So in a sense, we're picking on them because, you know, they're the strongest. They're an interesting case. They're a good test case here. And I think as we look to the future of what software will be, certainly for the stuff that we use at Amp, if there's something that our agent is the primary user of, or where we've ever been annoyed at the UI or their salespeople wanting us to get on the phone with them or anything about that. That's danger territory. And I think a lot of companies are seeing that. Like, if you are not, if you're using a tiny fraction of the features of a software package, it is so tempting to just go and build it yourself. So think your dashboard application. And in that case, a lot of the complexity in that is like their own programming language on top. And you You're not even interacting with that. You get zero value from that anymore. It also is kind of annoying to know that you're— maybe this is not the right way to think about it, but you're paying for a massive feature that suddenly you no longer use.

29:03Thorsten BallYeah.

29:04Quinn SlackAnd it's really tempting. And I do think that people will be doing that more and more, but there's this view. A lot of people talk about like, oh, you know, vibe-coded Salesforce from scratch, all of that, that people were talking about that last year. I think that's dumb because It's not Salesforce that you would vibe code from scratch. It's the 2% feature set of Salesforce that you need.

29:27Thorsten BallYou're not gonna vibe code Excel, but you're gonna vibe code that one spreadsheet that gets sent around and that kind of, I don't, I wanted to use the word load-bearing again, that's load-bearing for your one workflow. Like that's what you're gonna vibe code or build it yourself. Yeah.

29:45Quinn SlackYeah. there could be a core where someone, maybe at your company, maybe someone externally has built the really good primitives that would let you then vibe code. Because vibe coding, if that means starting from scratch every single time, even as models are getting better and better, you want consistency. And take like Granola, the recording app for Mac. I, in doing dictation on AMP, I've run into a lot of macOS APIs around audio. And there's a lot of footguns. There's just a lot of complexity that models still do not get it right regularly. So you want someone who's figured out all of that, but then you could configure your Jellyware granola to route it here and to understand these dictation things, to not send all your data to the other company, all these things. So I think it's, you want to do it on top of a, a core that's good.

30:35Thorsten BallYeah.

30:36Quinn SlackSo it's not that there's no role for software packages that are made by someone else. It's that, that is completely changing that rule.

30:45Thorsten BallYes, I 100% agree. Like, even with agents, I do not want to think about authentication. I do not. I just don't want to do it. It's like, what I want is I have an app idea and authentication should just be there. Like, somebody else should think about it. And that's not the agent. That's not a thing that the agent can do. You need management thinking. You need thinking of different options and weighing the trade-offs and whatnot. I don't wanna think about payments. I don't wanna think about— when I'm building an application, I don't wanna think about the storage layer, you know, like the database or whatever. So—

31:21Quinn SlackYeah. And you're talking about when you're building on top of a core, you want someone to have thought about that so that you don't have to think about it.

31:28Thorsten BallYeah, yeah, yeah, exactly. And man, like people say like vibe coding is fun. No, no, no. The most fun with programming, and we can say this is what we do with AMP, the most fun with agents and programming is When there is a foundation that's solid and that's working, and you just can let the agent go loose on it and change the UI or different things or add new features or remove features and whatnot without having to bump into those foundations. Like, if you have an app where authentication is set up and the secrets are set up and your production deployment is set up and you just have to tell the agent, deploy the new version, Then it's fun. But like the, it's always been the case that the worst part of programming is this whole wiring that shit together and putting the env vars in here and the secrets over there and this, do this. And oh, the SDK is not configured correctly because it's the YAML and not the env var and blah. Like all of the glue code, right? It's the worst part of programming. But now with agents, when people talk about vibe coding, that's kind of the shit that's left for you to do. Because the agent gets to have the fun and write the code. And now you have to do the fucking secrets management and the deployment and this stuff.

32:38Quinn SlackYeah. And a lot of the times the agent can't do that because Google's OAuth flow will block bots.

32:43Thorsten BallYeah. Yeah. It's like, oh, well, the agent can paint on the canvas now. All you have to do is go and buy the fucking paint. But what I want is somebody to bring the paint and the brushes, and then I can tell the agent, go and do this. Yeah.

32:57Quinn SlackAnd that's what we have when we are building and using AMP. And a lot of the joy, just the sheer wonder and awe that we get to experience is because we get to build on this foundation. And also, we get so much feedback. We have tons of users who give us ideas. And why don't you share the realization we had?

33:19Thorsten BallYeah. The realization is that there's 2 AMPs. One AMP is the one that our customers use. And by all accounts, it's a great product. They love it. They have fun with it. But then, the AMP we use internally is the AMP where I can wake up on a Tuesday and think, I want my orbs to be colored in the sidebar, and send off an agent. And then, I can say, yeah, let's put this behind a feature flag and ship it. And then, I post it on our Slack channel. What do you guys think about this? Then, I tweet it out, and then, I get 10 people saying, I want this, and then, I can enable it. From idea to deployed version, it's really short. Like, the biggest blocker right now, I would say, is our deployment pipeline. That's still because our scale is now large. So, it takes a while to roll that stuff out safely to a bunch of these processes running.

34:17Quinn SlackUsually, like 10 minutes, though. Well, yeah, there's some stuff for other stuff.

34:23Thorsten BallFor other stuff, like example yesterday, like again, like shout out to Ian. He was like, hey, can you build this for me? I sent off the prompt, it did it, and then it was like my 8:30. I had to put the kids to bed and whatnot. And I was like, I really want to get like— I really want to tell him it's done, but then I was like, okay, it's 8:30, gotta put the kids to bed. I'm going to bed then. Like, I I don't want to roll this. That was the biggest thing, but it was done, done. It was not done, done, but it was done. So, yeah, that's the realization that AMP for us is what we previously called Jellyware, right? It's high quality, it's high taste, and I'm quoting here, right? I'm not bragging. And yet, we don't develop in the traditional sense of how other people develop software. There's a lot of freedom and a lot of trust in this team on how and when people ship stuff. And it's not vibe-coded. Like, it's not slop.

35:22Quinn SlackAnd— But it's like when we use AMP, we are like homeowners. And with homeowners, you can make your house your own. You can put whatever on the walls. You can paint it however you want. You got to do all the maintenance. And if you see something that's wrong, no one else is going to fix it. You got to go fix it yourself. And that's how we feel. And there's a certain joy in that responsibility. And then the other way of using software is where you're just an end user and you cannot go change it when you're renting an apartment.

35:51Thorsten BallYeah.

35:52Quinn SlackAnd it's nice that you can call the landlord and they can go fix the washer and dryer and you don't have to deal with that. But hey, there's a lot of stuff you'd like to change about it and you can't. And there's pros and cons to both. But all of a sudden, it's become so much easier to go and change the house and so much cheaper. And now, a lot more people understandably want to be homeowners. And so, we're thinking about, for us, how we can get that AMP way to more people.

36:22Thorsten BallYeah.

36:23Quinn SlackAnd how we can give that to you in AMP and also how we can help you get that in more of the software that you're using. And it does require change in how companies work. Like if we, instead of building AMP the way that we do, which is small team, a ton of ownership, the ability to fix things, to change things, not a lot of process. If instead we had a very slow issue product management roadmap process, and then we had a very slow code review and deployment and approval process, Well, then it wouldn't matter that we have the ability to change stuff about AMP because we just wouldn't be able to. We'd be bottlenecked on either side.

37:03Thorsten BallYeah.

37:04Quinn SlackSo, we want to, you know, as one thing, export the AMP way to more teams, share more about what that means and why the joy of being able to push to main and why that can lead to higher quality and what you need to do differently on your team to make it so you can do that in a high-quality way.

37:22Thorsten BallYeah, that's what lies beyond the orbs, right? Like what's the next steps? Like that's, yeah, like orbs are just the next, it's the next stepping stone, right? To fully leaning into this. And it's funny, by the way—

37:38Quinn SlackSomeone who has terrible process on either side, for them, if it takes you a week to ship something and 3 weeks to figure out what to ship, then an EC2 cloud with a bunch of shell scripts is gonna be just as good as AMP. Neither of them are gonna be that good. AMP really shines when you can move so fast. That's when the multiplayer stuff, the portals, the orbs, the ability to do so much at once, the ability to have it monitor into production because it pushed and it is rolling it out. All of that stuff, we built AMP for that. So, you know, also us exporting the AMP way will help more people see why orbs are different.

38:15Thorsten BallIt's only because you just brought it up, I just realized how seldom I now think about these companies with traditional processes, where we get into contact with them when they said, can you make Orbs work for us in this workflow or in this environment? And man, I guess every software company should throw away, even as a thought experiment, how they've done things. and then restart from scratch and think like, okay, if we take the trajectories that we just talked about, prices of intelligence will go down, speed of tokens will go up, the intelligence itself will get smarter. If we take these trajectories and take them as a given, how would we develop software? And I think the outcomes would look completely different. That's the summary of this conversation, right? The funny thing is, the interesting thing to me is when we see companies, we mentioned Linear, 2 others come to mind like Raycast, Shopify, for example, they're leaning heavily into— they were founded pre-2025, even maybe 2020. I don't know with Raycast, but like pre-agentic companies and they're leaning heavily into AI. And so, you see them reshaping the company to this. But the weird thing here is that it's selection bias because we don't ever hear about the companies that don't reinvent themselves. They just die a quiet death. I'm not calling out a name, but I found a company. I posted it now on Slack. Maybe I could say it's about email, but it was like, The website looked like 20, you know, 2004, like 2008 maybe. And it's like, that's not like— who are you selecting? You're selecting for people that use their computer like it's 2008. And you— it's not like they will go out with a big bang. They will go out with a whimper because the world has moved on and it's just not going to be interesting. So I don't know where I'm going with this. I'm just saying, like, you need to rethink how you develop—

40:28Quinn SlackWell, it's easier said than done.

40:29Thorsten BallYeah, it's easier said than done, sure.

40:31Quinn SlackJust saying to everyone in the world, go rethink everything. But that's why as we are thinking about, given that we have AMP, how can we have the biggest impact? And to be clear, it's not about us suddenly making AMP bend over backwards to accommodate these old workflows. No, we want to evangelize this new way of working, and also take the fact that a lot of the software that these companies use is— you can now replace it with something that has 10% of the features, no settings screen, is maybe 1% of the cost, and is infinitely more customizable.

41:08Thorsten BallYeah.

41:08Quinn SlackAnd you don't have to be a company that is ready to be agent-pilled. You don't have to be someone who's using the agents directly in all the ways that we might want you to in order to get a ton of value from that. Because you can go and say, I used to have this CRM where, you know, one really smart person on my team had to manage it basically, you know, all the time. It was super clunky, it was super slow. And now, I can just say, hey, I want you to connect to Meta and pull in the data from here for our ad campaign. And that is something that anyone can appreciate.

41:45Thorsten BallYeah.

41:46Quinn SlackIt's how can we get this impact out more? And that's why I don't think that on the margin, the 100th person building their own agent platform product, I don't think that's going to move the needle. I think it's time for us at AMP and really software developers to think about how you can actually go solve the problem, not build software to build software to solve the problem.

42:08Thorsten BallYeah. And you just reminded me of something or made it click. Which is when you know in that triangle of build versus buy versus build and buy. I don't know what the, but there's a thing here where what you just said, right? Like it's easier said than done rethinking your business constantly while the models are getting better every two months, right? And while the costs are changing, and while the the how good they can do different things is constantly changing. So. This is what we're doing at AMP. Like, if you're using AMP and paying us, you're literally paying us to constantly rethink how software is being done. And this is what I do all day long, right? And that's in the product. Like, the way we think about how software can be built nowadays, that's the product. That's an expression. And it will look different in 2 months, right?

43:00Quinn SlackYeah.

43:01Thorsten BallSo, that's another thing. So, in times of these technological changes, It's not just like a static build versus buy because the landscape is completely changing, right? It's like, do I make my own rain jacket or do I buy a rain jacket? No, no, no. It's like you're going in the jungle and there's like a guide there and he's going to give you the clothes for the day because he can read the weather there. It doesn't matter whether you buy or build your own rain jacket. The weather's going to change. And you want the guy in there who knows, like, can read the clouds and whatnot, right? Not to oversell it, but I think that's—

43:37Quinn SlackYeah.

43:37Thorsten BallI don't know what that analogy is, but, you know, like AI land is a jungle.

43:42Quinn SlackWell, I think if you import that to software, it's— in the past, software was boring. It had to be mediocre because of that principle of the software business. You build it in year 1 and then you milk it for the next 10 years by—

43:54Thorsten BallYes.

43:54Quinn SlackBecause it's zero marginal cost copying. You know, you want it to appeal to the broadest market possible. And now that's changing. Like, why do I use Ghosty? Because I trust Mitchell in—

44:07Thorsten BallYeah.

44:07Quinn SlackWhat he thinks a terminal emulator should be. I want to see what is the issue tracker not made by a company that has to water it down to appeal to a lot of people, but what is the issue tracker made by someone that I really respect that has crazy ideas, who's going to show me not just, oh, it works, but how I should be doing this in the future or how we do video calls, how we plan our team events, how we do travel, all these. And so, in the past, if you had a great idea here, you would have to go and go hire a bunch of people, spend all your time doing that, not on the product. You'd go raise VC money.

44:47Thorsten BallYeah.

44:47Quinn SlackAnd in the end, the incentives would not be to build the most appealing opinionated correct product and there would only be like a few competitors. Now, there can be so many more competitors that you could even have, I want to use so-and-so famous developers or famous product person's product because their name is what gives me trust in it being different. And they can make so much more money and it can be cheaper because there's so much less overhead that they need to go and build that and distribute that. and sell that, they do not need to make a settings screen. They do not need to make it multi-tenant. They do not need to do all of this other stuff that they used to have to do. And that is really exciting. So, in the same way that we want AMP to be a representation of how we think you should be building, and if you trust us, then use AMP and we think you'll have a great time. I want that for all the software that I use.

45:37Thorsten BallYes.

45:38Quinn SlackYou know? It's—

45:40Thorsten Ballyou're paying somebody else to think about the problem constantly. Like, that's what you want. I want Mitchell out there thinking about terminals constantly. I want whoever the person is at Google Meet who thought of when I switch tabs, they show me that little popup, you know, where the Google Meet meeting goes on. Great job. Like, I want you to think about this. I don't wanna think about it. Which goes back to your point of like, you want to do the thing and not build the thing to do the thing. And yes, intelligence is cheap and tokens are cheap. And does labor, like intelligence labor, is cheap, but the thinking that a human does about a problem space is still different because it's continuous and they have different options. They weigh different options. They're human. So, that's kind of what's worth it. That's what you're paying for. I want somebody— I want like a Rasmus Andersson-designed text editor or whatever because I don't want to think through the edge cases or Zed or whatever. I can whip up my own, but for certain things, I want other people to think about it. Yeah, but yeah, I think we should we should any I'll end it with a short anecdote because you brought up Ghosty and the terminal. So we're all we all killed our local dev. I don't know, barely use it anymore, right? And I posted an outside. I said like in the last six weeks, I forgot all of the jujitsu or JJ shortcuts that I set up the aliases. I was completely lost. And then Tim Culverhouse replied. And Till Coverhaus, for those who don't know, he built his own terminal emulator. He built his own, I think, Wayland compositor window manager. He built his own email client. He built like everything he uses, he built himself. His own shell. He has his own shell running in his own terminal. And he replied saying, I have the best terminal and the best shell in the world and I don't use them. And it's, and it's sad smiley because he's fully in orbs now. So yeah, what does that say about, you know, the value of software?

47:38Quinn SlackYeah. And everything is changing faster. So if you were someone who has thought X and then you found X to not be true, I mean, we have, and it's very humbling, then you just get your mind to a place where you permanently are more open of crazy changes out there.

47:55Thorsten BallYeah. Wild times. All right, we'll see you next time.

0:00Quinn SlackJust building software to do x is only valuable if you can get people to to pay you, and I don't I think that that's gonna be a lot harder given that people can just make the software themselves. So developers need to think of themselves as doing x, not building software to do x. Yeah. And I think that's gonna be a rude shift because if anything, software developers on average have shifted to be more general, more horizontal in what they're building. The richest companies on earth are pouring

0:26Thorsten Ballall of their cash into this.

0:28Thorsten BallAnd we're not even seeing the results of this yet. Like, this is still being built. If you now think,

0:35Thorsten Balloh, I I have to be, stingy

0:38Thorsten Ballwith my monthly subscription, and I cannot run out of token. I think in a year or two, you will find that the tokens are everywhere. It's like electricity. It's like you're not measuring the watts coming out of your wall. And I think it's the same thing with tokens. There will be so many tokens everywhere

0:53Thorsten Balland so fast

0:55Thorsten Ballthat people won't recognize computing anymore.

1:06Thorsten BallAnd we're back. Another episode of raising an agent. This time, we're back home.

1:10Thorsten BallOne of us in in San Francisco in Mill Valley, the other one

1:14Thorsten Ballin this really small town in Germany near Frankfurt Airport.

1:19Thorsten BallAnd, yeah, we're excited to talk again. Right? Like, Munich sparked a lot of ideas. Munich got us got the the stuff flowing in our heads, the jelly flowing.

1:29Quinn SlackYeah. We've been shipping a lot, and it's been a few weeks since Munich. And we had said then that the way that we build had changed more in the prior, like, four or six weeks than it had changed all of last year. And

1:42Quinn SlackI think we're still feeling that, and we're starting to think about looking even

1:46Quinn Slackfurther

1:47Quinn Slackahead. And what is the world gonna be like? And how can we take what we're building, what we're seeing,

1:53Quinn Slackand

1:54Quinn Slackget more people building in the AMP way, and also then thinking about what is that going to do to

2:00Quinn Slacksoftware and what that means for software. Right? Yeah. So Yeah. Tharsten,

2:06Quinn Slackpredictions.

2:07Quinn SlackWhat do you think is gonna be different six,

2:10Quinn Slacknine months from now?

2:12Thorsten BallI think

2:14Thorsten Ballall of the trajectories will continue.

2:16Thorsten BallLike, I don't think there's gonna be this rug pull that most people say there might be or there will be. So I think the price of intelligence will drop. And people are saying, well, the models are more expensive. That's not well, the newest ones. Right? But, like, the price of the old models has dropped considerably.

2:34Thorsten BallAnd I think in, say, half a year,

2:38Thorsten Ballthe price of g p d five six levels of intelligence will be dirt cheap compared to right now.

2:43Thorsten BallAnd, you know, the Fable, incredibly expensive. I think intelligence at that level will be cheap.

2:50Thorsten BallI also think if you look at the trajectory of tokens per second, right, OpenAI is rolling out their super fast mode.

2:59Thorsten BallI don't have access. I wish I had access. It's it's what is it? 750 tokens per second, which is

3:06Thorsten Ballmind blowing. Had access, we would not be able to say. So I yeah. If I had access, I wouldn't be here talking, to be honest. Like, I I I wouldn't

3:14Thorsten Ballsend token.

3:15Thorsten BallIt that's mind blowing. So I think

3:18Thorsten Ballthis is

3:19Thorsten BallI'm not sure if this is correct, actually, but I think it's close to, like, the build out we're seeing with AI, the data centers, the GPUs.

3:27Thorsten BallIt's one of the biggest infrastructure build outs in the history of humanity.

3:31Thorsten BallYou know? Like, there's the richest companies on earth are pouring all of their cash into this.

3:38Thorsten BallAnd we only just we're not even seeing the results of this yet. Like, this is still being built.

3:43Thorsten BallSo

3:45Thorsten Ballif you now think,

3:47Thorsten Balloh, I I have to be, stingy with my monthly subscription,

3:51Thorsten Balland I cannot run out of tokens, and I count the tokens, and, you know, 50 k tokens is,

3:57Thorsten Ballsomething that I measure.

3:58Thorsten BallI think in a year or two or maybe three, you will find that the tokens are everywhere. It's like electricity.

4:04Thorsten BallIt's like you're not measuring the watts coming out of your wall, and you won't measure you know, you measure them in fucking megawatts and kilowatts,

4:13Thorsten Balland you don't care about, like,

4:16Thorsten Ball50 watts. Right? And I think it's the same thing with tokens. There will be so many tokens everywhere and so fast

4:23Thorsten Ballthat people won't recognize computing anymore. I think

4:27Thorsten Ballbased on these projections,

4:29Thorsten Balla lot of

4:31Thorsten Ballthings that we think about how software development

4:34Thorsten Ballis

4:35Thorsten Ballwill change.

4:37Thorsten BallWell, do you disagree with any of these predictions, or do you wanna add some?

4:41Quinn SlackI agree.

4:42Quinn SlackIntelligence is getting cheaper and cheaper.

4:45Quinn SlackAnd anyone who thinks there's gonna be a rug pull, the, quote, unquote, subsidies will end, There's no evidence that subscriptions,

4:53Quinn Slackwhich is what people usually mean, are actually

4:56Quinn Slackunprofitable,

4:56Quinn Slackand it's hard to actually tease it apart in the same way that it's

5:00Quinn Slacknot like United Airlines is going to stop having economy seeding

5:04Quinn Slackbecause that's been a subsidy, and they're gonna pull the subsidy away. No. You know, once there's a big upfront expense, it makes sense to spread that in a price discriminated way across the largest pot possible population of buyers.

5:15Quinn SlackYeah. And, also, totally agree that if you are in a role where you're having to be really stingy with the intelligence that you can use, then that's a signal. In much the same way, if you're someone who's getting paid

5:30Quinn Slackone twentieth of what a San Francisco engineer makes, then find a way to get to San Francisco or closer to San Francisco. It's a signal that what you're doing is not benefiting enough from tokens. You're not creating enough, value. You're not in that good position. But if you're thinking that, then you're probably the kind of person who can create a ton of value if you get into a job or other setting where

5:52Quinn Slackthe tokens can flow.

5:54Quinn SlackAnd at the same time, there's a lot of noise going on right now. First,

6:00Quinn Slackit feels like

6:01Quinn Slackeverybody is

6:03Quinn Slacktrying to build the same thing. Now, you know, Thurston, you and I have been building DevTools

6:07Quinn Slackfor a long time. And for the longest time,

6:12Quinn Slackeveryone said, oh, you can't make any money in it. And Yeah. Now people have completely changed to the point where it feels like

6:20Quinn Slackeverybody is building dev tools, and that also

6:23Quinn Slackis probably not correct.

6:26Quinn SlackAnd, ultimately, software,

6:29Quinn Slackit has to solve the problem. You know? Just just building software to do x

6:35Quinn Slackis only valuable if you can get people to to pay you, and I don't I think that that's gonna be a lot harder given that people can just make the software themselves. So I think more and more, we've said this before, but developers need to think of themselves as doing x, not building software to do x.

6:50Quinn SlackYeah. And I think that's gonna be a rude shift, because if anything, software developers on average have shifted to be more general, more horizontal in what they're building. I do have a little pet theory that I wanna offer here, and it's just it's a pet theory. Say the inflection point was November 25, right, which

7:07Thorsten Ballthis is my other pet theory. It wasn't for us. I don't think November 25 was that big. I don't think there was a bump in the curve. I think it was a curve.

7:16Quinn SlackExcept for Gemini.

7:17Thorsten BallRemember? That was well, let's not talk about it. But I

7:21Thorsten Ballthink

7:23Thorsten BallI think it was a smooth curve, but there's a lot of engineers who've kind of woken up to agents now. And now we've seen, like, in the last few months,

7:33Thorsten Ballpeople and I've said this before

7:35Thorsten Ballon this podcast,

7:36Thorsten Ballbut it's it's even more than this where people are now

7:40Thorsten Balllike, they're hacking on their own agents and, like, the plug in. And it's I thought this would go away, but it's still a thing that's going on where people build their own tooling around this stuff and blah blah blah blah blah. And my pet theory is engineers have this thing

7:56Thorsten Ballnow where

7:58Thorsten Ballthey they go, it can't be this easy. Right? Like, you know, like, how we always try to fix our tools and build, like, some elaborate workflows and automate stuff? Like, I mean, not me. I'm the guy who copy and pastes you know, I'm I'm not doing this. But, like, if you send a prompt to Fable I did this yesterday. I used Fable. I'm kinda on a

8:17Thorsten Balltwo day Fable binge. I don't know. Maybe it's because I wanna spend money after we can put Jatchapati subs into AMP, and now I'm like, well, you know, gotta

8:26Thorsten Ballnot, you know, gotta use Fable. So I sent a prompt yesterday. Ian Lanceman, shout out to Ian. He sent a tweet. He's like, I wanna be able to set the default branch in AMP for orbs.

8:36Thorsten BallSo I sent a prompt. I tweeted the prompt. It was like a text message I sent I would send to somebody.

8:42Thorsten BallAnd Fable went and did it, and it just added a feature. It didn't add complexity. It kind of wired that new field, the default branch through, like, our different services. I had it watch deployment. I had it make sure it rolled out. I tested it, and it works.

8:56Thorsten BallI do think, and here's the pet theory, that engineers then go and look at this and go, oh, it can't be this easy. Right? How about I

9:04Thorsten Balladd a bunch of stuff? I add, like, some tools and, like, some shortcuts and, like, a review thing and this process and that process and blah blah blah. Because it doesn't appeal to the engineers of the last say twenty years that you just say something in plain English

9:20Thorsten Balland it goes and does it. You know? Like, I think there's, like, something like this going on where you wanna kinda add, like, some value and that's that's how you build, like, these other engineering the the this whole elaborate stuff around it. But Yeah. Most of if you're a furniture builder, you would probably feel ashamed if someone comes over your house and it's all IKEA furniture.

9:40Thorsten BallAnd when you sit down on that IKEA couch, you're gonna hear every little creak and it's gonna Yeah. Yeah. Hell out of you. I mean, we all know the story. Right? That some guy posted, I don't know, twenty years ago on the Internet. They they were a three d designer

9:53Thorsten Balland basically, they figured out that every time they had like something they made and they showed it to their boss and said like, does this look good? The boss would always say

10:02Thorsten Balljust

10:03Thorsten Ballto the boss would feel they have to add value so they would ask, like, change this even though it was completely nonsensical.

10:10Thorsten BallSo the designer added a little duck to everything

10:14Thorsten Balland then the boss would always go, that looks great, but get rid of the duck. That they'll get and then they would feel happy about it. They added something. So I do think there's there might be something going on here where engineers feel like it can't be as easy as just saying something to the model and build software. Right? And

10:31Thorsten Ballso they have to add something.

10:33Quinn SlackYeah. But yeah. So with an an agent, though, you wrote the how to build an agent that

10:38Quinn Slackturned on a lot of people to how simple the idea of an agent loop with tools could be,

10:44Quinn Slackand

10:45Quinn Slackthat was

10:47Quinn Slackpretty cool. That opened people's eyes. That's a new primitive.

10:51Quinn SlackWhat about all these people who say,

10:53Quinn Slackoh, Orbs, how is that different from just EC two and you set up a VPN and you have some orchestration?

11:01Quinn SlackYou know? Like,

11:02Quinn Slackwhat what are they gonna run into that they can't build themselves?

11:06Thorsten BallThere's multiple things going on. I think well, first of all, we have to admit that building stuff is pretty easy now getting started with stuff. Right? And we had like the first version of what's now called orbs

11:21Thorsten Ballin two weeks, a week or something.

11:23Thorsten BallAnd then we ran into the long tail with that infrastructure

11:28Thorsten Ballprovider we used and

11:30Thorsten Ballof issues and in a whole

11:33Thorsten Ballslog. And then I I think I rewrote it on the new platform over the course of a weekend. And this was not me sitting here for twelve hours. This was me starting proms and whatnot.

11:42Thorsten BallAnd so admittedly,

11:44Thorsten Ballit is easy to build a lot of stuff. But when

11:48Thorsten Ballsomebody says,

11:49Thorsten Balloh, I built orbs and it's like easy to and a VPN and or, you know, I have, like, 15 Tmax sessions running on my home lab and, like, 18 different clock codes in it. And you can get far, but it's obviously not the same experience. Like, the the value of orbs is that you have

12:07Thorsten Ballall of the things in the same

12:09Thorsten Ballpackage. You know? Like, it's not you don't have to manage TMac sessions. The beauty of Orpz is you don't have to think about the machine. You just send your prompt. You get the results back. You preview it in a portal. You can share it with your colleagues and whatnot. Like, you don't have to worry about how the sausage is made. And

12:26Thorsten Ballthere's also people who are offended by this where they're like, I wanna see how the sausage is made, you know, but I don't know. I don't know what to say. Like, it's

12:34Thorsten Ballthe same thing maybe. Like, are you really concerned about how we spun up these sandboxes or VMs? Like, do you what do you wanna see? You know? Like, you can specify hooks and scripts now. You can define your own batch. It would pretty exposing

12:46Thorsten Balla bunch of stuff, but I I think it's underselling the whole experience. Like, you can get, like, 80% of the way there, but what we found is

12:54Thorsten Ballthe 20%

12:55Thorsten Ballmakes a lot of difference. It's weird how hard it was to get this nailed down because

13:01Thorsten Ballit's not like super hard programming. Also, what does that mean anymore? I'm not going on this sidetrack, but, you know, like, you'd asked me five years ago, super hard program would be like an optimizing pass for

13:14Thorsten Ballan optimizing compiler. You know? I'm

13:16Quinn Slackpretty sure that has kind of flipped now. Like Yeah. It's like the design. And Yeah. The design all these pieces fit together.

13:24Thorsten BallYeah. Into a whole. And, like, the product, what are you building?

13:29Thorsten BallIn what sequence? When are you building it? What are the trade offs? You know, all of this stuff. So when we

13:37Thorsten Ballbuilt this, we've this was really hard. Like, how what are the primitives, for example? And what are the things that the invariants of the product? One thing that I think we got right is

13:49Thorsten Ballyou start an orb, you get a single URL, you get a fresh orb for new threads. That was controversial when I said this. Right? Like, everybody was like, no. No. No. I wanna spawn multiple agents in an orb, you know, in the same thing. And now look at us, like,

14:03Thorsten Ballthere would be different product where you can spawn multiple agent. Like, you know, like, with different and select the specific orb and whatnot. And that would be a different thing. So

14:11Thorsten Ballpeople can go far and wire their own stuff together, but wiring your own stuff together is not building the same product. And there's a lot of, like, decisions

14:22Thorsten Ballthat are not code level that go into making the full thing,

14:26Thorsten Balland those seem to be valuable. Right? If you look at the feedback we get where people say it's it just works or it's a great experience or,

14:34Thorsten Ballsomebody tweeted this just an hour ago, like,

14:38Thorsten Ballthe your UI or your, like, your front end is completely on a different level and whatnot. Right? And,

14:45Thorsten Ballobviously, like, we have great designers and engineers on the team who who really sweat over the details here. And

14:51Thorsten BallI don't know. I don't know. I I don't know how you can replicate this with your homemade VPN,

14:56Thorsten Balleasy to

14:58Quinn Slackthe TNX thing. I think What do you think? Two is to look at it. One is if you're just trying to have fun or just build something

15:05Quinn Slackfor yourself to use

15:08Quinn Slackand you don't need to

15:10Quinn Slackmake it work at, like, team scale, you don't need to get other humans involved, you don't need to deal with the complexities that come with other people

15:17Quinn Slackalso wanting to make it work their way, then, you know, you could probably build something

15:22Quinn Slacklike a Yeah. Agent platform that's pretty fun.

15:25Quinn SlackYeah. But once you start making it so that an entire team can work on it, that all the different ways they use it work, that you can read other people's threads, you can have the, you know, puck like orchestrator

15:38Quinn Slackagent work across everything in it, that all the concepts fit together.

15:43Quinn SlackThat's hard. And the hard part for us has been no single person on our team. You know, our team is 20 people. All of AMP is 20 people. No single person on our team has every single piece of the system and the design in their head. And so much of it is figuring out how we can have all the different pieces that we all know and feel really strongly about work well together.

16:06Quinn SlackAnd

16:07Quinn Slackthat is tough. We all use the product all the time. We have a small team that's highly trusted. So we've optimized AMP for being able to do that. But that's where if it's you know you don't really have that need if you're one person building it for yourself. And if you're one person building it for your whole company, you're gonna run into all the same problems that we have. And then Yeah. People are gonna be saying, well, why can't I use AMP or, you know, all the other? Now there's 87 different things

16:33Quinn Slackthat are in this category, and that is probably gonna give you a headache. But the more important thing is, like, what is it's fun. I don't get me wrong. That's part of why we do it, but you're not actually

16:46Quinn Slackcreating

16:47Quinn Slackvalue.

16:48Quinn SlackIt's not like any of these agents like AMP. It's not like we're charging you a ton for the joy of using our product. We're making money. We're profitable, but you're not gonna save a lot of money. You're probably gonna well, actually lose money on net because there's lot of efficiency stuff we do. So what is the business value you're you're doing? You can't be making software to do x. You most certainly cannot be making software to make software to do x. You gotta be doing x these days. Yeah. Yeah. But, I mean, it points at a at a real tension. Right? The whole build versus buy thing.

17:20Thorsten BallRight? Where should we buy something

17:23Thorsten Balland not build it ourselves? And

17:26Thorsten Balltwo years ago,

17:28Thorsten Ballyou know, besides

17:30Thorsten Balloperational costs and like guarantees

17:33Thorsten Balland support and whatnot,

17:35Thorsten Ballyou had the big thing of I don't wanna build this myself or I can build it myself.

17:40Thorsten BallBut now in the age of I take a screenshot and I say, you know, build this for me,

17:46Thorsten Ballthat equation has changed. Like, that scale

17:49Thorsten Ballhas been tipped. And I think the interesting bit is that the models are now

17:54Thorsten Ballweighed on the scale, on the build versus buy scale, and and the weight of the models changes constantly. Like, it's it's you know, like, you constantly have to reevaluate.

18:03Thorsten BallAnd, I mean, admittedly, I build a backtracker into

18:07Thorsten BallAMP last week. Right? And it was

18:10Thorsten Ballone shot at, and it was deeply integrated into our system.

18:13Thorsten BallSo

18:15Thorsten Ballthat might have been even easier than spending the time to setting up linear or something because now the value is in deeply integrating stuff, which is why we're even talking. Right? Like, this whole thing is is incredibly fascinating to me that if you look at the history of software, and we've said this a thousand times, software was hired and expensive to write, slow.

18:34Thorsten BallBasically, you want the investment you make when writing software to pay off or have a return,

18:42Thorsten Ballso you generalize

18:43Thorsten Ballthe software that you write, so you can sell it or reuse it multiple times in different use cases. Right? You don't build one off software.

18:52Quinn SlackThat's not doesn't pay or it hasn't paid. But The beauty of a software business is you build it, and then for the next ten years, your customers will stay on it, and you can sell it,

19:03Thorsten Balland infinite copies could be made at zero marginal cost, and you can make money on each of those. So you recoup your fixed expense. Yeah. Even even if you're an agency and you build customer specific things, most agencies have then developed some kind of framework or, like, some I guess you can call it a stock or whatever, like, something that they reuse across projects. Right? Which means they can save time, like a common admin interface or whatever.

19:27Thorsten BallOr,

19:29Thorsten Ballthink of, like, WordPress. Right? Which is I mean, it's been ten years since I checked. But back then, it was most configurable pieces of software. Like, the admin interface had a lot of menus.

19:39Thorsten BallAnd

19:40Thorsten Ballnow it's like, well, you could take a screenshot or four of different websites and have like one shot a thing

19:47Thorsten Balllike your hair salon website or like your whatever restaurant website.

19:53Thorsten BallAnd then every time you wanna make a change,

19:55Thorsten Ballyou can ask the agent again to do it. Right? So this has changed. And and and like what I find fascinating is that a lot of practices

20:04Thorsten Balland methodologies

20:05Thorsten Ballof software,

20:07Thorsten Ballwe still do them or

20:09Thorsten Ballor they kinda crumbling,

20:11Thorsten Ballbut a lot of stuff doesn't hold anymore. Right? Where

20:14Quinn SlackYeah. We talked about Before you even get onto that, I think you mentioned this, but I just wanna put it to you again. Basically,

20:20Quinn Slackwhat we are saying is we see a lot of people say, oh, AMP or other agents, I could just replace that with EC two and some shell scripts and get the same thing. And and we

20:31Quinn Slackunderstand that, but we don't think that that's the right thing for them to do. But at the same time, you're saying, I was able to code

20:38Quinn Slackthe parts of linear, an issue tracker that I need from scratch.

20:41Quinn SlackI, you know, quote, unquote, replaced linear with a few CRUD forms.

20:46Quinn SlackSo which is correct? How could we say that in one case, you know, AMP is worth it? But in the other case, no, linear is not worth it. What what makes it different?

20:56Thorsten BallThat's a good question. I think

20:59Thorsten Ballwhat's really different about

21:03Thorsten Balllet's start on this end. I would say the linear or bug report case

21:07Thorsten Ballis so different because

21:10Thorsten Ballthe workflow around how you do bugs

21:13Thorsten Ballis different in each company.

21:15Thorsten BallAnd it's like the it's it's

21:18Thorsten Balllike bug tracking or issue reporting or

21:22Thorsten Ballknowledge management, like, it's different in every company. That's why Jira has 15,000 options. Right? And why every Jira is different. It's not because

21:30Thorsten Ballpeople don't like the button, but it's because they they work in a different way.

21:35Thorsten BallI think Could you say the way that people use agents and do software development is different in every company and therefore every company I guess you could. I'm starting to I'm I'm thinking about it, but I would argue that the it's like the thing I said at the start. Right? Like, you have, like you need to specify what you want and the model does the thing, and there's not a lot of this

21:55Thorsten Ballit's handed over to the model is, I guess, what I'm saying. There's not a lot of touch points anymore with the humans. Maybe that's a difference. In that,

22:02Thorsten Ballyou know, like, previously, like, your deaf environment had, like, a lot of touch points where the human were the humans were changing things. But now, man, like, in orbs, like, they just whip up software and tear it down, and I don't care. And

22:16Quinn SlackI don't know. What do you think? Like, this is this is interesting. What do you think? I think this is the big question that everyone has about software is what software is below the waterline

22:25Quinn Slackwhere Yeah. You don't wanna go dive and mess with it, and what software is above the waterline where you wanna, you know, go go build?

22:33Quinn SlackYeah. If you just look at people's behavior,

22:36Quinn SlackI think you've seen a lot of people shift from

22:40Quinn Slackwanting to spend their time in issue land

22:44Quinn Slackand project land and typing that in, and now they want to spend their time in agent land. And you cannot argue with people's revealed preferences or actual actions. And so there's something about,

22:55Quinn Slackwell, why do we think AMP or, generally, the agent is different from the issue tracker? It's because we have found it to be

23:03Quinn Slackincredibly

23:04Quinn Slackproductive

23:05Quinn Slackto work there, far more than issues

23:08Quinn Slackever would. And so there's there's more

23:11Thorsten Ballvalue to having that be really good and solid. Oh, I see what you're saying. So you okay. I see what you're saying. So the value the business value of the software has changed, and you would say that building your own for low value things

23:27Thorsten Ballis

23:28Quinn Slackthat's that's you know? We don't spend that much time in it ourselves.

23:32Quinn SlackSo Yeah. There's not that much complexity that we need to replicate. So in the past, maybe we would have used

23:39Quinn Slack60% of the features of

23:41Quinn SlackLinear.

23:42Quinn SlackAnd Linear is an awesome, awesome product.

23:45Quinn SlackYou know, we we've used it, and it's the best of its category.

23:50Quinn SlackBut now we might only be using 10%. And I think there's a lot of people that are using some of Linear's agent features, and they like that, but we would only be using

23:58Quinn Slack10% of it. So I get I think there's also an element where

24:02Quinn Slacklinear was built for a workflow that is undergoing a lot of change.

24:06Quinn SlackAnd if you have I think there's also something fundamental about software businesses

24:11Quinn Slackthat

24:12Quinn Slackwere started before 2025

24:14Quinn Slackand have not completely reinvented

24:17Quinn Slackthemselves and

24:19Quinn Slackburned the bridges. I think there's also a lot of valid skepticism that that those workflows will not be

24:25Quinn Slackbe there.

24:27Quinn SlackI think there's also something about if, like, an agent is at the core that then that just makes the software so much more pliable and moldable. Like, in the way with AMP, if you wanna do something, most of the time, you can just ask Puck to do it. If you want an integration,

24:40Quinn Slackyou can just ask, you know, Puck or AMP in a thread to go set up a webhook.

24:44Quinn SlackAnd

24:45Quinn Slackyeah. I don't know. But, you know, this this is what we're thinking about right now. Yeah. Another angle is the whole

24:51Thorsten BallI don't really like the term, but, like, the whole system of record type of thing. Meaning, like, some software is more important than others. I think that's kinda what people wanna say when they use the term.

25:01Thorsten BallMeaning,

25:02Thorsten Ballwith orbs,

25:03Thorsten Ballyou you're spending money on VMs. So as soon as you're spending money, like, significant amount of money, that changes this whole equation. Right? And you have the tokens flowing through the system. You wanna keep track of this. I do think, like, the

25:18Thorsten Ballwhole process of that touches the code is important. So, you know, like, there's something systemy of recordishy

25:25Thorsten Ballabout, like, the orbs.

25:27Thorsten BallWhereas with the tickets thing that I built, it's not. Right? Like, for us, that's not a cost. Like, it's it's it's a crud app in our admin

25:35Thorsten Ballcrud app. It's basically free,

25:37Thorsten Balland that changes it for me. And, like, the whole Orb stuff, that's our main application. That's the main thing that's that's load bearing to use the

25:47Thorsten Ballstolen term, the agent stole from us. But I don't know. Maybe that's this. But I do think, like, what you said is very interesting with you have the agent at the core, and you can change it. And

25:58Thorsten Ballthere's

25:59Thorsten Ballobviously now this tension of Linea was built in a time where you had to for different customers,

26:07Thorsten Ballthey they have different workflows. So you had to have the option to custom labels or like custom workflows or like custom boards and whatnot.

26:15Thorsten BallAnd admittedly,

26:17Thorsten Balllinear was, I think, three years ago, they were pretty minimal. Right? And they said like, we only have this. And then people said, like, to follow I don't know. That's not true. But, like, couple years back, they started adding more features and it became more complex

26:30Thorsten Ballcomplex.

26:31Thorsten BallAnd, obviously, if you're if you're SaaS software,

26:34Thorsten Ballyou have to be this complex because you have multiple customers and they give you different amounts of money and they have different usage and different workflows. So you your software has to be generic for different things. Right? But

26:46Thorsten Ballwith agents now,

26:48Thorsten Ballif you build your own,

26:51Thorsten Ballit doesn't have to be generic at all. And that kinda changes the whole cost thing of the software again. Like, it makes it less complex and whatnot. But obviously, like, what I'm talking about here is like this triangle

27:03Thorsten Ballof,

27:04Thorsten Ballwell, you know, you can buy the SaaS thing and it's complex in the back end and a lot of work to build because you have to generalize it or you build your own and then it's kinda maybe achievable, but it's customized to workflow. But then on the other end of the triangle,

27:19Thorsten Ballthere's this thing that, you know, with orbs and whatnot where it's it's load bearing, it costs money.

27:25Thorsten BallYou do wanna modify it, but

27:29Thorsten Ballyeah. I don't know. Like, I think everybody's trying to figure out where what goes where on this triangle.

27:34Quinn SlackYeah. And with AMP, we we

27:37Quinn Slackkill features that we no longer think are good. So there's not cruft in AMP that we're keeping around because

27:43Quinn Slacksome customer nine months ago paid us for a year, and they'd be pissed if we take it out. No. I mean, everyone knows with AMP that there's no cruft in it. And by the way, Linear

27:53Quinn Slackis of basically any software company best positioned here, and they've done

27:58Quinn Slackway better than

28:00Quinn Slackpretty much any other company I know in in staying relevant. So in a sense, we're picking on them because, you know, they're the strongest. They're an interesting case. They're a good test case here.

28:08Quinn SlackAnd I think if as we look to the future of what software will be,

28:14Quinn Slackcertainly for the stuff that we use at AMP,

28:17Quinn Slackif there's something that our agent is the primary user of or where we've ever been annoyed at the UI

28:25Quinn Slackor

28:26Quinn Slackthere are salespeople wanting us to get on the phone with them or anything about that, that's danger territory, and I think a lot of companies are are seeing that. Like, if you are not if you're using a tiny fraction of the features of a software package,

28:40Quinn Slackit is so tempting to just go and build it yourself.

28:44Quinn SlackSo think your dashboard application.

28:47Quinn SlackAnd in that case, a lot of the complexity in that is, like, their own programming language on top, and you're not even interacting with that. You get zero value from that anymore. It also is kind of annoying to know that you're maybe this is not the right way to think about it, but you're paying for a massive feature that suddenly you no longer use.

29:04Quinn SlackAnd it's really

29:06Quinn Slacktempting. And I do think that people will be doing that more and more, but there's this view and a lot of people talk about, like, oh, you know, Vibe coded Salesforce from scratch,

29:15Quinn Slackall of that. That people were talking about that last year. I think that's dumb because it's not Salesforce

29:20Quinn Slackthat you would Vibe code from scratch. It's the 2%

29:24Quinn Slackfeature set

29:25Quinn Slackof Salesforce

29:26Thorsten Ballthat you need. Gonna you're not gonna vibe code Excel,

29:30Thorsten Ballbut you're gonna vibe code that one spreadsheet

29:33Thorsten Ballthat gets sent around and that kinda

29:36Thorsten BallI don't I wanted to use the word load bearing again. That's load bearing for your one workflow. Like, that's what you're gonna vibe code or build it yourself. Yeah. Yeah. Or there could be a core where someone, maybe your company, maybe someone externally has built the

29:52Quinn Slackreally good primitives that would let you then vibe code. Because vibe coding, if that means starting from scratch every single time, even as models are getting better and better, you want consistency.

30:04Quinn SlackAnd take, like, granola,

30:07Quinn Slackthe recording app for a Mac. I in doing dictation,

30:11Quinn Slackon AMP, I've run into a lot of Mac OS APIs around audio, and there's a lot of foot guns. There's just a lot of complexity. The models still do not get it right regularly. So you want someone who's figured out all of that, but then you could configure your jellyware granola

30:25Quinn Slackto route it here and to understand these dictation things, to not send all your data to the other company, all these things. So I think it's you wanna do it on top of a a core that's good. So it's not that there's no role for software packages that are made by someone else. It's that Yeah. That is completely

30:43Quinn Slackchanging

30:44Quinn Slackthat role.

30:46Thorsten BallI I 100%

30:48Thorsten Ballof it. Like, with agents,

30:50Thorsten BallI do not want to think about authentication.

30:52Thorsten BallI do not I just don't wanna do it. It's like what I want is

30:57Thorsten BallI have an app idea and

30:59Thorsten Ballauthentication

31:00Thorsten Ballshould just be there. Like, else should think about it. And that's not the agent. That's not the thing that the agent can do. You need management thinking. You need thinking of different options and weighing the trade offs and whatnot. I don't wanna think about payments.

31:14Thorsten BallI don't wanna think about when I'm building an application, I don't wanna think about the storage layer, you know, like the database or whatever. So Yeah. And you're talking about when you're building on top of a core. You want someone to have thought about that so that you don't have to think about it. And, man, like, people say, like, vibe coding is fun. No. No. No. The the most fun with programming

31:34Thorsten Balland we'll we can say this is what we do with AMP. The most fun with agents and programming is

31:39Thorsten Ballwhen they're

31:40Thorsten Ballthere is a foundation that's solid and that's working, and you just can let the agent go loose on it and and change the UI or different things or add new features or remove features or whatnot without having to bump into those foundations. Like, if you have an app where authentication is set up and the secrets are set up and, like, your production

32:00Thorsten Balldeployment is set up and you just have to tell the agent deploy the new version, then it's fun. But it's like the it's always been the case that the worst part of programming is this whole wiring that shit together and putting the in force in here and the secrets over there and this do this and, oh, the SDK is not configured correctly because it's the YAML and not the and blah. Like, all of the glue code. Right? It's the worst part of programming. But now, with agents, when people talk about vibe coding, that's kind of the shit that's left for you to do because the agent gets to do all of the gets to have the fun and write the code and now you have to do the fucking secrets management and the deployment and this stuff. Yeah. And a lot of times the agent can't do that because Google's OAuth flow will block bots. Yeah. Yeah. It's like, oh, well, the agent can paint on the canvas now. All you have to do is go and buy the fucking paint. But what I want is to somebody bring the paint and the brushes, and then I can tell the agent, go and do this. You know? Yeah. And And that's what we have when we are building and using AMP. Yes. And a lot of the joy,

33:04Quinn Slackjust the sheer wonder and awe that we get to experience is because we get to build on this foundation.

33:10Quinn SlackAnd, also, we get so much feedback. We have tons of users who Yeah. Give us ideas.

33:15Quinn SlackAnd

33:16Quinn Slackwe

33:17Quinn Slackwhy don't you share the realization we had?

33:20Thorsten BallYeah. The realization is that there's two amps.

33:23Thorsten BallAnd

33:24Thorsten Ballone amp is the one that our customers use.

33:28Thorsten BallAnd

33:29Thorsten Ballby all accounts, it's a great product. They love it. They they have fun with it.

33:34Thorsten BallBut then the amp we use internally is the amp where I can wake up on a Tuesday and think I want my orbs to be colored in the sidebar and send a threat send off an agent,

33:46Thorsten Balland then I can say, yeah, let's put this behind a feature flag and ship it, and then I post in our Slack channel. What do you guys think about this? Then I tweet it out, and then I get 10 people saying I want this, and then I can enable it. It's it's

33:59Thorsten Ballfrom idea to deployed version, it's really

34:02Thorsten Ballit's short. Like, the biggest blocker right now, I would say, is our deployment pipeline. That's still because we have our scale is now

34:10Thorsten Balllarge, So it takes a while to roll that stuff out safely to a bunch of these

34:16Thorsten Ballprocesses running.

34:17Quinn SlackUsually, like, ten minutes, though.

34:20Quinn SlackWell,

34:21Thorsten Ballthere's Yeah. There's some stuff For other stuff. For other stuff. Like, I'll Yeah. Example, yes. And I like, again, like, shout out to Ian. He was like, hey, can you build this for me? I sent off the prompt. It did it. And then it was like my 08:30.

34:34Thorsten BallI had to put the kids to bed and whatnot. And I was like,

34:37Thorsten BallI really wanna get like, I really wanna tell him it's done.

34:41Thorsten BallBut then I was like, okay. It's 08:30.

34:43Thorsten BallGotta put the kids to bed. I'm going to bed then. Like, I don't wanna roll this, you know, like, that was the biggest thing, but it was done done. You know? Like, it was not done done, but it was done. So, yeah, that's that's a realization that AMP for us is what we previously called jellyware. Right? It's it's high quality.

35:01Thorsten BallIt's

35:02Thorsten Ballhigh tastes,

35:04Thorsten Balland I'm quoting here. Right? I'm not bragging.

35:06Thorsten BallAnd and yet, we don't develop in a traditional sense of how other people develop software. Like, there's a lot of freedom and a lot of trust in this team on how and when people ship stuff. And it's not vibe coded. Like, it's not it's not slop.

35:22Quinn SlackAnd But it's like when we use AMP, we are like homeowners.

35:26Quinn SlackAnd with homeowners, you can make your house your own. You can put whatever on the walls. You can paint it however you want. You gotta do all the maintenance. And if you see something that's wrong, no one else is gonna fix it. You gotta go fix it yourself. And that's how we feel. And there's a certain joy in that responsibility.

35:44Quinn SlackAnd then the other way of using software is where you're just an end user, and you cannot go change it when you're, you know, renting an apartment.

35:52Quinn SlackYeah. And it's nice that you can call the landlord, and they can go fix the washer and dryer, and you don't have to deal with that. But, hey. You know, there's a lot of stuff you'd like to change about it, and you can't. And there's pros and cons to both,

36:04Quinn Slackbut all of a sudden, it's become so much easier to go and change the house and so much cheaper.

36:10Quinn SlackYeah. And now, you know, a lot more people understandably wanna be homeowners. And so we're thinking about, for us, how we can get that AMP way

36:20Quinn Slackto more

36:21Quinn Slackpeople

36:23Quinn SlackYeah. And how we can give that to you in AMP, and also how we can help you get that in more of the software

36:29Quinn Slackthat you're using.

36:31Quinn SlackAnd it does require change in how companies work.

36:34Quinn SlackLike, if we

36:37Quinn Slackinstead of building AMP the way that we do, which is small team, a ton of ownership, the ability to fix things, to to change things, not a lot of process. If instead we had a very slow issue product management road map process,

36:51Quinn Slackand then we had a very slow code review and deployment and approval process,

36:56Quinn Slackwell, then

36:57Quinn Slackit wouldn't matter that we have the ability to change stuff about AMP because we we just wouldn't be able to. We'd be bottlenecked on either side.

37:04Thorsten BallSo Yeah. We wanna,

37:06Quinn Slackyou know, as one thing, export the AMP way

37:08Quinn Slackto more teams. Share more about what that means and why

37:12Quinn Slackthe joy of being able to push to main and and why that can lead to higher quality and what you need to do differently on your team to make it so you can do that in a high quality way. Yeah.

37:23Thorsten BallThat's

37:25Thorsten Ballwhat lies beyond the orbs. Right? Like, what's the next steps? Like, that's, yeah. Like, orbs are just the next it's the next stepping stone, right, to to fully leaning into this. And

37:37Thorsten Ballit's funny because someone

37:39Quinn Slackwho has terrible process on either side, for them,

37:43Quinn Slackif it takes you a week to ship something and three weeks to figure out what to ship, then an e c two cloud with a bunch of shell scripts is gonna be just as good as AMP. Neither of them are they gonna be that good. AMP really shines when you can move so fast. That's when the multiplayer stuff, the portals, the orbs, the ability to do so much at once, the ability to have it monitor into production because it pushed, and it is rolling it out. All of that stuff, we built AMP for that. So Yeah. You know? Also, us exporting the AMP way will help more people see why orbs are different.

38:15Thorsten BallIt's only

38:17Thorsten Ballbecause you just brought it up. I just realized, like, how seldom I now think about these companies with traditional processes.

38:24Thorsten BallYou know, where we get into contact with them when they said, like, can you make Orbs work for us with

38:31Thorsten Ballin this workflow, in this environment?

38:33Thorsten BallAnd, man, like, I guess, like, every software company should throw away,

38:39Thorsten Balleven as a thought experiment how they've done things and then restart from scratch and think like, okay. If we take the trajectories that we just talked about, intelligence will prices of intelligence will go down, speed of tokens will go up, The intelligence itself will get smarter.

38:55Thorsten BallIf we take these trajectories

38:57Thorsten Balland take them as a given, how would we develop software?

39:00Thorsten BallAnd I think a lot the outcomes would look completely different. That's the whole that's the summary of this conversation. Right? But

39:07Thorsten Ballthe funny thing is the interesting thing to me is when we see companies,

39:12Thorsten Ballyou met you know, we mentioned linear.

39:15Thorsten BallTwo others come to mind like Raycast,

39:18Thorsten BallShopify, for example. They're leaning heavily into they were pre

39:22Thorsten Ballfounded pre 2025,

39:24Thorsten Balleven maybe 2020

39:26Thorsten Ballin the I don't know with Raycast. But, like,

39:29Thorsten Ballpre agentic companies,

39:31Thorsten Balland they're leaning heavily into AI.

39:33Thorsten BallAnd so you see them how reshaping

39:36Thorsten Ballthe company to this.

39:38Thorsten BallBut the the weird

39:40Thorsten Ballthing here is that,

39:42Thorsten Ballyou know, it's selection bias because we don't ever hear about the companies that don't reinvent themselves. Like, they just die a quiet death.

39:50Thorsten BallLike, I'm not calling out a name, but I found a company. I posted it now. It's like

39:56Thorsten Ballmaybe I could say it's about email, but it was like the website looked like

40:01Thorsten Ball'20, you know, 2004,

40:04Thorsten Balllike 2008

40:05Thorsten Ballmaybe. And

40:07Thorsten Ballit's like, that's not like, who are you selecting? You're selecting for people that use their computer like it's 2008

40:14Thorsten Balland you

40:17Thorsten Ballit's not like they will go out with a big bang. They will go out with a whimper because the world has moved on and it's just not gonna be interesting. So I don't know where I'm going with this. I'm just saying like, you need to rethink how it's yourself. Done.

40:30Thorsten BallYes. These are said everything. In the world.

40:32Quinn SlackGo rethink everything.

40:34Quinn SlackBut Yeah. That's true. Why as we are

40:37Quinn Slackthinking about

40:38Quinn Slackgiven that we have AMP, how can we have the biggest impact?

40:43Quinn SlackAnd to be clear, it's not about us

40:45Quinn Slacksuddenly making AMP bend over backwards to accommodate these old workflows. No. We want to evangelize

40:51Quinn Slackthis new way of working and also

40:54Quinn Slacktake the fact that a lot of the software that these companies use is

40:59Quinn Slackyou can now

41:00Quinn Slackreplace it with something that has 10% of the features, no setting screen,

41:05Quinn Slackis maybe 1% of the cost, and is infinitely more customizable.

41:09Quinn SlackAnd you don't have to be a company that is ready to be agent pilled. You don't have to be someone who's using the agents directly in all the ways that we might want you to in order to get a ton of value from that because you can go and say,

41:23Quinn SlackI used to have this CRM where, you know, one really smart person in my team had to manage it basically, you you know, all the time. It was super clunky. It was super slow.

41:33Quinn SlackAnd

41:33Quinn Slacknow I can just say, hey. I want you to connect to Meta and pull in the data from here for our ad campaign. And that is something

41:42Quinn Slackthat anyone can appreciate.

41:46Quinn SlackYeah. So it's how can we get this impact out more? And that's why I don't think that on the margin, the, you know, hundredth person building their own agent platform product,

41:57Quinn SlackI don't think that's gonna move the needle. I think it's time for us at AMP and really software developers to think about how you can actually go solve the problem, not build software to build software to to solve the problem or something like that.

42:11Thorsten BallAnd you just reminded me of something or made it click, which is

42:14Thorsten Ballwhen you know, in that triangle of build versus buy versus build and buy, I don't know what the

42:20Thorsten Ballbut there's a thing here where what you just said. Right? Like, it's easier said than done rethinking your business constantly.

42:28Thorsten BallWhile the models are getting better every two months. Right? And while the costs are changing, and while the the the the how good they can do different things is constantly changing. So

42:40Thorsten Ballthis is what we're doing at AMP. Like, if you're using AMP and paying us, you're literally paying us to constantly rethink how software is being done. Like and this is what I do all day long. Right? Like and that's in the product. Like, the way we think about how software can be built nowadays, that's the product. That's an expression. And it will it will look different in two months. Right? So Yeah.

43:01Thorsten BallThat's another thing. So in times of these technological

43:04Thorsten Ballchanges,

43:05Thorsten Ballit's not just like a static build versus buy because the landscape is completely changing. Right? It's like,

43:12Thorsten Ballyou know, do I make my own rain jacket or do I buy a rain jacket? No. No. No. It's like you're going in the jungle and there's like a guy there and he's gonna give you the clothes for the day because he can read the weather there. It it it doesn't matter whether you buy or build your own rain jacket. The weather's gonna change

43:30Thorsten Balland you want the guy in there who knows like can read the clouds and whatnot. Right? Not to oversell it, but I think that's Yeah. I don't know what that analogy is, but, you

43:41Quinn Slackknow, like, AI land is a I think if you import that to software, it's in the past, software was boring. It had to be mediocre.

43:48Quinn SlackBecause of that principle of the software business, you build it in year one, and then you milk it for the next ten years by Yes. Because it's zero marginal cost copying, you know, you want it to appeal to the broadest market possible.

44:00Quinn SlackAnd now that's

44:02Quinn Slackchanging.

44:03Quinn SlackLike, why do I use Ghosty? Because I trust Mitchell in Yeah. What he thinks a terminal emulator

44:09Quinn Slackshould be.

44:11Quinn SlackI wanna see what is the issue tracker

44:14Quinn Slacknot made by a company that has to water it down to appeal to a lot of people, but what is the issue tracker made by someone that I really respect that has crazy ideas who's gonna show

44:24Quinn Slackme not Yeah. Just, oh, it works, but how I should be doing this in the future.

44:30Quinn SlackOr how we do

44:33Quinn Slackvideo calls, how we plan our team events, how we do travel, all these. And so in the past, if you had a great idea here, you would have to go and go hire a bunch of people, spend all your time doing that, not on the product. You'd go raise VC money. And in the end, the incentives would not be to build the most opinionated,

44:50Quinn Slackcorrect product,

44:52Quinn Slackand there would only be, like, a few competitors now. There could be so many more competitors that you could even have, I want to use so and so famous developers or famous product person's product because their name is what gives me trust in it being different. And they can make so much more money, and it could be cheaper because there's so much less overhead that they need to go and build that and distribute that and sell that.

45:17Quinn SlackThey do not need to make a setting screen. They do not need to make it multitenant. They do not need to do all of this other stuff that they used to have to do. And that is really exciting. So in the same way that we want AMP to be a representation of how we think you should be building. And if you trust us, then use AMP, and we think you'll have a great time. I want that for all the software that I use. Yes. You know? It's

45:40Thorsten BallYou're paying somebody else to think about the problem constantly. Like, that's what you want. I want Mitchell out there thinking about terminals constantly.

45:48Thorsten BallI want whoever the person is at Google Meet who thought of when I switch tabs, they show me that little pop up. You know, where the Google Meet meeting goes on, great job. Like, I want you to think about this. I don't wanna think about it, which goes back to your point of, like, you want to do the thing and not build the thing to do the thing. And, yes, intelligence is cheap and tokens are cheap and thus labor,

46:09Thorsten Balllike intelligence labor

46:11Thorsten Ballis cheap,

46:12Thorsten Ballbut, like, the thinking that a human does about a problem space is still different. Right? Because it's continuous,

46:18Thorsten Balland they have different options, a way different options than a human. So

46:22Thorsten Ballthat's kind of what you what what's

46:25Thorsten Ballworth it. Right? That's what you're paying for. I want somebody I want, like, Erasmus Anderson design text editor or whatever because I don't wanna think through the edge cases or zed or whatever. Like, I can whip up my own, but for certain things, want other people to think about it.

46:40Thorsten BallYeah. But, yeah, I think we should we should any I'll I'll end it with a short anecdote because you brought up Ghosty and the terminal. So we're all we all killed our local dev

46:50Thorsten BallI don't know. Barely use it anymore. Right? And I posted in our site. I said, like, in the last six weeks, I forgot all of the jujitsu

46:58Thorsten Ballor JJ shortcuts that I set up the aliases. I was completely lost. And then Tim Culverhouse replied.

47:05Thorsten BallAnd Tim Culverhouse,

47:06Thorsten Ballfor those who don't know,

47:08Thorsten Ballhe built

47:10Thorsten Ballhis own terminal emulator. He built his own, I think, Wayland Compositor, Window Manager. He built his own email client. He built like, everything he uses, he built himself. He owns Shell. He has its own Shell running in his own terminal. And he replied saying, I have my I have the best terminal and the best Shell in the world, and I don't use them.

47:30Thorsten BallAnd it's and a sad smiley because he's fully in Orbs now.

47:34Thorsten BallSo Yeah. What does that say about, you know, the value of software?

47:39Quinn SlackYeah. And everything is changing faster. So if you were someone who has thought x and then you found x to not be true, I mean, we have, and it's very humbling,

47:49Quinn Slackthen you just get your mind to a place where you permanently are more open of crazy changes

47:54Quinn Slackout there.

47:56Thorsten BallYeah. Wild times. Alright.

47:59Thorsten BallWe'll see you next time.

0:00Just building software to do X is only valuable if you can get people to pay you. And I think that

0:06that's going to be a lot harder given that people can just make the software themselves. So developers

0:11need to think of themselves as doing X, not building software to do X. And I think that's

0:16going to be a rude shift because if anything, software developers on average have shifted to

0:20be more general, more horizontal in what they're building.

0:23The richest companies on earth are pouring all of their cash into this.

0:28and we're not even seeing the results of this yet like this is still being built if you now think

0:34oh i i have to be uh stingy with my monthly subscription and i cannot run out of tokens i

0:41think in a year or two you will find that the tokens are everywhere it's like electricity it's

0:46like you're not measuring the watts coming out of your wall and i think it's the same thing with

0:50tokens there will be so many tokens everywhere and so fast that people won't recognize computing

0:57anymore.

1:05And we're back.

1:07Another episode of Raising an Agent.

1:08This time we're back home.

1:10One of us in San Francisco in Mill Valley,

1:12the other one in this really small town in Germany near Frankfurt

1:17Airport.

1:19And yeah, we're excited to talk again, right?

1:22Munich sparked a lot of ideas.

1:24Munich got us, got the stuff flowing in our heads,

1:27the jelly flowing. Yeah, we've been shipping a lot. And it's been a few weeks since Munich. And

1:33we had said then that the way that we build had changed more in the prior like four or six weeks

1:39than it had changed all of last year. And I think we're still feeling that. And we're starting to

1:44think about looking even further ahead. And what is the world going to be like? And how can we take

1:51what we're building, what we're seeing, and get more people building in the AMP way.

1:57And also then thinking about what is that going to do to software and what that means for software, right?

2:03Yeah.

2:04So, Thurston, predictions.

2:07What do you think is going to be different six, nine months from now?

2:12I think all of the trajectories will continue.

2:16I don't think there's going to be this rug pull that most people say there might be or there will be.

2:22So I think the price of intelligence will drop.

2:26And people are saying, well, the models are more expensive.

2:28That's not, well, the newest ones, right?

2:30But the price of the old models has dropped considerably.

2:34And I think in, say, half a year, the price of GPT-5, 6 levels of intelligence will be dirt cheap compared to right now.

2:43and you know the fable incredibly expensive i think intelligence at that level will be cheap

2:50i also think if you look at the trajectory of tokens per second right uh open ai is rolling

2:56out their super fast mode um i don't have access i wish i had access it's it's what is it 750 tokens

3:03per second which is mind if we had access we would not be able to say so i if i had access i

3:10wouldn't be here talking to be honest i wouldn't send token it that's mind-blowing so i think

3:17this is i'm not sure if this is correct actually but i think it's close to like the build-out we're

3:24seeing with ai that data centers to gpus it's one of the biggest infrastructure build-outs in the

3:30history of humanity you know like there's the richest companies on earth are pouring all of

3:36their cash into this and we only just we're not even seeing the results of this yet like this is

3:42still being built so if you now think oh i i have to be uh stingy with my monthly subscription and i

3:51cannot run out of tokens and i count the tokens and you know 50k tokens is uh something that i

3:57measure i think in a year or two or maybe three you will find that the tokens are everywhere it's

4:03like electricity it's like you're not measuring the watts coming out of your wall and you won't

4:08measure you know you measure them in fucking megawatts and kilowatts and you don't care about

4:14like 50 watts right and i think it's the same thing with tokens there will be so many tokens

4:20everywhere and so fast that people won't recognize computing anymore i think based on these

4:28projections, a lot of things that we think about how software development is, will change.

4:37Do you disagree with any of these predictions or do you want to add some?

4:41I agree. Intelligence is getting cheaper and cheaper. And anyone who thinks there's going

4:47to be a rug pull, the quote unquote subsidies will end. There's no evidence that subscriptions,

4:53which is what people usually mean, are actually unprofitable. And it's hard to actually tease it

4:58apart in the same way that it's not like United Airlines is going to stop having economy seeding

5:03because that's been a subsidy and they're going to pull the subsidy away. No, once there's a big

5:08upfront expense, it makes sense to spread that in a price discriminated way across the largest

5:13possible population of buyers. And also totally agree that if you are in a role where you're

5:20having to be really stingy with the intelligence that you can use, then that's a signal in much

5:26the same way if you're someone who's getting paid 120th of what a San Francisco engineer makes,

5:33then find a way to get to San Francisco or closer to San Francisco. It's a signal that

5:37what you're doing is not benefiting enough from tokens. You're not creating enough value. You're

5:43not in that good position. But if you're thinking that, then you're probably the kind of person

5:46who can create a ton of value if you get into a job or other setting where the tokens can flow.

5:53and at the same time there's a lot of noise going on right now first it feels like everybody is

6:02trying to build the same thing now you know thurston you and i have been building dev tools

6:07for a long time and for the longest time everyone said oh you can't make any money in it and now

6:15people have completely changed to the point where it feels like everybody is building dev tools and

6:22that also is probably not correct. And ultimately software, it has to solve the problem. You know,

6:31just, just building software to do X is only valuable if you can get people to pay you. And

6:37I don't, I think that that's going to be a lot harder given that people can just make the software

6:42themselves. So I think more and more, we've said this before, but developers need to think of

6:47themselves as doing X, not building software to do X.

6:50And I think that's going to be a rude shift because if anything,

6:53software developers on average have shifted to be more general,

6:56more horizontal in what they're building.

6:58I do have a little pet theory that I want to offer here.

7:01And it's a pet theory.

7:03Say the inflection point was November 25, right?

7:06Which this is my other pet theory.

7:09It wasn't for us.

7:10I don't think November 25 was that big.

7:12I don't think there was a bump in the curve.

7:14I think it was a curve.

7:16Except for Gemini, remember?

7:19Well, let's not talk about it.

7:20But I think it was a smooth curve.

7:24But there's a lot of engineers who've kind of woken up to agents now.

7:29And now we've seen like in the last few months, people – and I've said this before on this podcast.

7:36But it's even more than this where people are now like they're hacking on their own agents and like the plug-in.

7:43And I thought this would go away, but it's still a thing that's going on where people build their own tooling around this stuff and blah, blah, blah, blah, blah.

7:52And my pet theory is engineers have this thing now where they go, it can't be this easy, right?

8:01You know how we always try to fix our tools and build some elaborate workflows and automate stuff?

8:07I mean, not me.

8:08I'm the guy who copy and pastes.

8:09I'm not doing this.

8:10But like if you send a prompt to Fable, I did this yesterday.

8:14I used Fable.

8:14I'm kind of on a two-day Fable binge.

8:18I don't know.

8:19Maybe it's because I want to spend money after we can put ChatGPT subs into AMP.

8:23And now I'm like, well, you know, got to use Fable.

8:27So I sent a prompt yesterday.

8:29Ian Lanceman, shout out to Ian.

8:31He sent a tweet.

8:32He's like, I want to be able to set the default branch in AMP for Orbs.

8:36So I sent a prompt.

8:37I tweeted a prompt.

8:38It was like a text message I would send to somebody.

8:42Fable went and did it, and it just added a feature.

8:45It didn't add complexity.

8:46It kind of wired that new field, the default branch, through our different services.

8:51I had it watch deployment.

8:53I had it make sure it rolled out.

8:54I tested it, and it works.

8:56I do think, and here's the pet theory, that engineers then go and look at this and go,

9:01oh, it can't be this easy, right?

9:02How about I add a bunch of stuff?

9:05I add like some tools and like some shortcuts and like a review thing and this process and that process and blah, blah, blah.

9:12Because it doesn't appeal to the engineers of the last, say, 20 years that you just say something in plain English and it goes and does it.

9:22You know, like I think there's like something like this going on where you want to kind of add like some value.

9:27And that's how you build like these other engineering, this whole elaborate stuff around it.

9:32Well, it's like if you're a furniture builder, you would probably feel ashamed if someone comes over to your house and it's all IKEA furniture.

9:40And when you sit down on that IKEA couch, you're going to hear every little creak and it's going to go the hell out of you.

9:46I mean we all know the story, right?

9:48That some guy posted, I don't know, 20 years ago on the internet.

9:51They were a 3D designer and basically they figured out that every time they had like something they made and they showed it to their boss and said like, does this look good?

10:00But the boss would always say, the boss would feel they have to add value.

10:06So they would ask, like, change this, even though it was completely nonsensical.

10:10So the designer added a little duck to everything.

10:14And then the boss would always go, that looks great, but get rid of the duck.

10:18And then they would feel happy about it.

10:21They added something.

10:21So I do think there might be something going on here where engineers feel like it can't be as easy as just saying something to the model and build software, right?

10:30And so they have to add something.

10:33Yeah.

10:33So with an agent, though, you wrote the how to build an agent that turned on a lot of people to how simple the idea of an agent loop with tools could be.

10:44And that was pretty cool.

10:48That opened people's eyes.

10:49That's a new primitive.

10:50What about all these people who say, oh, Orbs, how is that different from just EC2 and you set up a VPN and you have some orchestration?

11:01Yeah.

11:02What are they going to run into that they can't build themselves?

11:06There's multiple things going on.

11:08I think – well, first of all, we have to admit that building stuff is pretty easy now getting started with stuff, right?

11:15And we had like the first version of what's now called Orbs in two weeks, a week or something.

11:23And then we ran into the long tail with that infrastructure provider we used and of issues and in a whole slog.

11:33And then I think I rewrote it on the new platform over the course of a weekend.

11:38And this was not me sitting here for 12 hours.

11:40This was me starting prompts and whatnot.

11:42And so admittedly, it is easy to build a lot of stuff.

11:46But when somebody says oh I built Orbs and it like EC2 and a VPN and or you know I have like 15 T sessions running on my home lab and like 18 different cloud codes in it And you can get far but it obviously not the same experience Like the value of Orbs is that you have all of the things in the same package

12:10You know, like it's not, you don't have to manage T-Mark sessions.

12:13The beauty of Orbs is you don't have to think about the machine.

12:16You just send your prompt, you get the results back, you preview it in a portal,

12:20you can share it with your colleagues and whatnot.

12:22And like, you don't have to worry about how the sausage is made.

12:25And there's also people who are offended by this,

12:28where they're like, I want to see how the sausage is made, you know?

12:31But I don't know.

12:32I don't know what to say.

12:33Like, it's the same thing maybe.

12:35Like, are you really concerned about how we spawn up these sandboxes or VMs?

12:39Like, do you, what do you want to see?

12:40You know, like you can specify hooks and scripts now.

12:43You can define your own bash.

12:44But we're pretty exposing a bunch of stuff.

12:47But I think it's underselling the whole experience.

12:49Like you can get like 80% of the way there, but what we found is the 20% makes a lot of difference.

12:57It's weird how hard it was to get this nail down because it's not like super hard programming.

13:04Also, what does that mean anymore?

13:06I'm not going on this sidetrack, but you know, like if you'd asked me five years ago, super hard program would be like an optimizing pass for an optimizing compiler.

13:15You know, I'm pretty sure that has kind of flipped now.

13:18like yeah it's like the design and yeah the design all these pieces fit together yeah into a whole

13:25and like the product what are you building in what sequence when are you building it what are

13:32the trade-offs you know all of this stuff so when we built this we've this was really hard like how

13:40what are the primitives for example and what are the things that the invariants of the product

13:46one thing that i think we got right is you start an orb you get a single url you get a fresh orb for

13:52new threads that was controversial when i said this right like everybody was like no no i want

13:56to spawn multiple agents in an orb you know in the same thing and now look at us like there would

14:03be different product where you can spawn multiple agents like you know like with different and select

14:08the specific orb and whatnot that would be a different thing so people can go far and wire

14:14their own stuff together, but wiring your own stuff together is not building the same product.

14:19And there's a lot of decisions that are not code level that go into making the full thing.

14:26And those seem to be valuable, right? If you look at the feedback we get where people say,

14:30it just works or it's a great experience or somebody tweeted this just an hour ago,

14:37like your UI or your front end is completely on a different level and whatnot, right?

14:43And obviously like we have great designers and engineers on the team who,

14:48who really sweat over the details here. And I don't know, I don't know.

14:52I don't know how you can replicate this with your homemade VPN,

14:56easy to the T-Nex thing.

14:59I think there's two ways to look at it.

15:02One is if you're just trying to have fun or just build something for yourself

15:06to use and you don't need to make it work at like team scale,

15:12You don't need to get other humans involved.

15:14You don't need to deal with the complexities that come with other people also wanting to make it work their way.

15:20Then, you know, you could probably build something like a agent platform.

15:24That's pretty fun.

15:26Yeah.

15:26But once you start making it so that an entire team can work on it, that all the different ways they use it work, that you can read other people's threads.

15:34You can have the, you know, puck like orchestrator agent work across everything in it, that all the concepts fit together.

15:42that's hard. And the hard part for us has been no single person on our team. You know, our team is

15:4920 people. All of AMP is 20 people. No single person on our team has every single piece of

15:55the system and the design in their head. And so much of it is figuring out how we can have all

16:01the different pieces that we all know and feel really strongly about work well together. And

16:06And that is tough.

16:08We all use the product all the time.

16:10We have a small team that's highly trusted.

16:13So we've optimized AMP for being able to do that.

16:15But that's where if it's, you know, you don't really have that need if you're one person building it for yourself.

16:20And if you're one person building it for your whole company, you're going to run into all the same problems that we have.

16:26And then people are going to be saying, well, why can't I use AMP or, you know, all the other.

16:31Now there's 87 different things that are in this category.

16:34And that is probably going to give you a headache.

16:36But the more important thing is like, what is, it's fun.

16:41I don't get me wrong.

16:42That's part of why we do it, but you're not actually creating value.

16:48It's not like any of these agents like AMP.

16:51It's not like we're charging you a ton for the joy of using our product.

16:56We're making money.

16:56We're profitable, but you're not going to save a lot of money.

16:59You're probably going to actually lose money on net because there's a lot of efficiency

17:02stuff we do.

17:03So what is the business value you're doing?

17:05You can't be making software to do X.

17:07You most certainly cannot be making software to make software to do X.

17:11You got to be doing X these days.

17:14Yeah.

17:14But I mean it points at a real tension, right?

17:18The whole build versus buy thing, right?

17:21Where should we buy something and not build it ourselves?

17:25And two years ago, besides operational costs and guarantees and support and whatnot, you had the big thing of, I don't want to build this myself or I can build it myself.

17:39But now in the age of I take a screenshot and I say, build this for me, that equation has changed.

17:47Like that scale has been tipped.

17:50And I think the interesting bit is that the models are now weight on the scale, on the build versus buy scale.

17:57And the weight of the models changes constantly.

18:00Like it's, you know, like you constantly have to reevaluate.

18:03And I mean, admittedly, I built a bug tracker into AMP last week, right?

18:08And it was one-shotted and it was deeply integrated into our system.

18:13So that might have been even easier than spending the time to setting up linear or something

18:19because now the value is in deeply integrating stuff, which is why we're even talking, right?

18:23Like this whole thing is incredibly fascinating to me that if you look at the history of software,

18:29and we've said this a thousand times, software was hard and expensive to write, slow.

18:34Basically, you want the investment you make when writing software to pay off or have a return.

18:41so you generalize the software that you write

18:44so you can sell it or reuse it multiple times

18:48in different use cases, right?

18:49You don't build one-off software.

18:52That doesn't pay or it hasn't paid.

18:55But the beauty of a software business is you build it

18:57and then for the next 10 years,

19:00your customers will stay on it and you can sell it

19:03and infinite copies can be made at zero marginal cost

19:06and you can make money on each of those.

19:08So you recoup your fixed expense.

19:10Yeah, even if you're an agency and you build customer-specific things, most agencies have then developed some kind of framework or like some, I guess you can call it a stock or whatever, like something that they reuse across projects, right?

19:23Which means they can save time, like a common admin interface or whatever.

19:27Or think of like WordPress, right?

19:30Which is, I mean, it's been 10 years since I checked, but back then it was most configurable pieces of software.

19:35Like the admin interface had a lot of menus.

19:38and now it's like well you could take a screenshot or four of different websites and have like

19:45one shot a thing like your hair salon website or like your whatever restaurant website and then

19:53every time you want to make a change you can ask the agent again to do it right so this has changed

19:59and and and like what i find fascinating is that a lot of practices and methodologies of software

20:06we still do them or they're kind of crumbling, but a lot of stuff doesn't hold anymore, right?

20:14Yeah. So before you even get onto that, I think you mentioned this, but I just want to

20:18put it to you again. Basically, what we are saying is we see a lot of people say, oh,

20:24AMP or other agents, I could just replace that with EC2 and some shell scripts and get the same

20:29thing. And we understand that, but we don't think that that's the right thing for them to do. But at

20:34the same time, you're saying I was able to code the parts of linear, an issue tracker that I need

20:40from scratch. I, you know, quote unquote, replaced linear with a few crud forms. So which is correct?

20:47How could we say that in one case, you know, AMP is worth it, but in the other case, no,

20:53linear is not worth it. What makes it different? That's a good question. I think what's really

21:00different about let's start on this end i would say the linear or buck report case

21:06is so different because the workflow around how you do bugs is different in each company

21:15and it's like the it's it's like buck tracking or issue reporting or knowledge management like

21:23it's different in every company that's why gyra has 15 000 options right and why every gyra is

21:28different it's not because people don't like the button but it's because they they work in a

21:33different way so i think you say the way that people use agents and do software development

21:38is different in every company and therefore every i mean i guess you could i'm starting i'm i'm

21:42thinking about it but i would argue that it's like the thing i said at the start right like

21:47you have like you need to specify what you want and the model does the thing and there's not a lot

21:54of this it's handed over to the model is i guess what i'm saying there's not a lot of touch points

21:59anymore with the humans maybe that's the difference in that you know like previously like your deaf

22:04environment had like a lot of touch points where the human were the humans were changing things but

22:09now man like in orbs like they just whip up software and tear it down and i don't care and

22:15i don't know what do you think like this is interesting what do you think i think this is

22:20the big question that everyone has about software. What software is below the waterline where you

22:26don't want to go dive and mess with it? And what software is above the waterline where you want to,

22:31you know, go, go build. If you just look at people's behavior, I think you've seen a lot

22:37of people shift from wanting to spend their time in issue land and project land and typing that in.

22:46And now they want to spend their time in agent land.

22:49And you cannot argue with people's revealed preferences, their actual actions.

22:54And so there's something about, well, why do we think AMP or generally the agent is different from the issue tracker?

23:00It's because we have found it to be incredibly productive to work there far more than issues ever would.

23:09And so there's more value to having that be really good and solid.

23:15Oh, I see what you're saying.

23:16So, okay, I see what you're saying. So the value, the business value of the software has changed. And you would say that building your own for low value things is that's, that's, you know, we don't spend that much time in it ourselves. So there's not that much complexity that we need to replicate. So in the past, maybe we would have used 60% of the features of linear and linear is an awesome, awesome product.

23:45We've used it and is the best of its category,

23:50but now we might only be using 10%.

23:53And I think there's a lot of people

23:54that are using some of linear agent features and they like that but we would only be using 10 of it So I think there also an element where linear was built for a workflow that is undergoing a lot of change

24:06And I think there's also something fundamental about software businesses

24:10that were started before 2025 and have not completely reinvented themselves

24:17and burned the bridges.

24:19I think there's also a lot of valid skepticism that those workflows

24:24will not be there.

24:26I think there's also something about

24:28if like an agent is at the core,

24:31that then that just makes the software

24:32so much more pliable and moldable.

24:34Like in the way with AMP,

24:35if you want to do something,

24:36most of the time you can just ask Puck to do it.

24:38If you want an integration,

24:40you can just ask Puck or AMP in a thread

24:43to go set up a webhook.

24:44And yeah, I don't know.

24:46But this is what we're thinking about right now.

24:48Yeah, another angle is the whole,

24:50I don't really like the term,

24:52but like the whole system of record type of thing,

24:54meaning like some software is more important than others.

24:57I think that's kind of what people want to say when they use the term,

25:01meaning with Orbs, you're spending money on VMs.

25:05So as soon as you're spending money, like significant amount of money,

25:09that changes this whole equation, right?

25:12And you have the tokens flowing through the system.

25:14You want to keep track of this.

25:16I do think like the whole process of that touches the code is important.

25:21So, you know, like there's something system-y of record-ish-y about like the orbs.

25:26Whereas with the tickets thing that I built, it's not, right?

25:30Like for us, that's not a cost.

25:32Like it's a CRUD app in our admin CRUD app.

25:35It's basically free.

25:37And that changes it for me.

25:39And like the whole orb stuff, that's our main application.

25:42That's the main thing.

25:43That's load-bearing to use the stolen term the agent stole from us.

25:49But I don't know, maybe that's this.

25:51But I do think like what you said is very interesting with you have the agent at the core and you can change it.

25:57And there's obviously now this tension of Linear was built in a time where you had to, for different customers, they have different workflows.

26:09So you had to have the option to custom labels or like custom workflows or like custom boards and whatnot.

26:15And admittedly, linear was, I think, three years ago, they were pretty minimal, right?

26:19And they said, like, we only have this.

26:21And then people said, like, I don't know, that's not true.

26:25But, like, a couple of years back, they started adding more features and it became more complex.

26:31And obviously, if you're SaaS software, you have to be this complex because you have multiple customers and they give you different amounts of money and they have different usage and different workflows.

26:41So your software has to be generic for different things, right?

26:45But with agents now, if you build your own, it doesn't have to be generic at all.

26:53And that kind of changes the whole cost thing of the software again.

26:56Like it makes it less complex and whatnot.

26:59But obviously, like what I'm talking about here is like this triangle of, well, you know, you can buy the SaaS thing and it's complex in the background.

27:08a lot of work to build because you have to generalize it or you build your own and then

27:13it's kind of maybe cheap to build but it's customized to workflow but then on the other

27:17end of the triangle there's this thing that you know with orbs and whatnot where it's it's load

27:23bearing it costs money you do want to modify it but yeah i don't know like i think everybody's

27:30trying to figure out where what goes where on this triangle yeah and with amp we we kill features

27:37that we no longer think are good.

27:39So there's not cruft in AMP that we're keeping around

27:42because some customer nine months ago

27:44paid us for a year and they'd be pissed if we take it out.

27:47No, I mean, everyone knows with AMP

27:49that there's no cruft in it.

27:51And by the way, Linear is of basically

27:54any software company best positioned here

27:56and they've done way better

27:58than pretty much any other company I know

28:01in staying relevant.

28:02So in a sense, we're picking on them

28:03because they're the strongest,

28:05they're an interesting case,

28:06they're a good test case here.

28:07And I think as we look to the future of what software will be, certainly for the stuff that we use at AMP, if there's something that our agent is the primary user of or where we've ever been annoyed at the UI or their salespeople wanting us to get on the phone with them or anything about that, that's danger territory.

28:31And I think a lot of companies are seeing that.

28:34Like if you are not, if you're using a tiny fraction of the features of a software package, it is so tempting to just go and build it yourself.

28:44So think your dashboard application.

28:47And in that case, a lot of the complexity in that is like their own programming language on top.

28:52And you're not even interacting with that.

28:53You get zero value from that anymore.

28:55It also is kind of annoying to know that you're – maybe this is not the right way to think about it, but you're paying for a massive feature that suddenly you no longer use.

29:04And it's really tempting, and I do think that people will be doing that more and more.

29:10But there's this view – and a lot of people talk about like, oh, VibeCoded Salesforce from scratch, all of that.

29:16People were talking about that last year.

29:17I think that's dumb because it's not Salesforce that you would VibeCode from scratch.

29:22It's the 2% feature set of Salesforce that you would need.

29:28You're not going to VibeCode Excel, but you're going to VibeCode that one spreadsheet that gets sent around and that kind of – I wanted to use the word load-bearing.

29:38That's load-bearing for your one workflow.

29:41Like that's what you're going to VibeCode or build it yourself.

29:44Yeah. Or there could be a core where someone, maybe your company, maybe someone externally has built the really good primitives that would let you then vibe code.

29:57Because vibe coding, if that means starting from scratch every single time, even as models are getting better and better, you want consistency.

30:04And take like Granola, the recording app for Mac.

30:09I, in doing dictation on AMP, I've run into a lot of macOS APIs around audio, and there's a lot of foot guns.

30:16There's just a lot of complexity that models still do not get it right regularly.

30:19So you want someone who's figured out all of that, but then you could configure your Jellyware granola to route it here and to understand these dictation things, to not send all your data to the other company, all these things.

30:32So I think it's you want to do it on top of a core that's good.

30:35So it's not that there's no role for software packages that are made by someone else.

30:41It's that that is completely changing that role.

30:47100%.

30:47Even with agents, I do not want to think about authentication.

30:52I do know.

30:53I just don't want to do it.

30:54It's like what I want is I have an app idea and authentication should just be there.

31:01Somebody else should think about it.

31:03And that's not the agent.

31:04That's not a thing that the agent can do. You need management thinking. You need thinking of different options and weighing the trade-offs and whatnot. I don't want to think about payments. I don't want to think about when I'm building an application, I don't want to think about the storage layer, like the database or whatever.

31:21So, yeah. And you're talking about when you're building on top of a core, you want someone to have thought about that.

31:26100%. Yeah, yeah, yeah. Exactly. And man, like people say like vibe coding is fun. No, no, no. The most fun with programming and we can say this is what we do with them. The most fun with agents and programming is when there is a foundation that's solid and that's working and you just can let the agent go loose on it and change the UI or different things or add new features or remove features and whatnot.

31:51without having to bump into those foundations.

31:53Like if you have an app where authentication is set up

31:57and the secrets are set up

31:59and like your production deployment is set up

32:02and you just have to tell the agent,

32:03deploy the new version, then it's fun.

32:05But it's always been the case

32:08that the worst part of programming

32:10is this whole wiring that shit together

32:12and putting the NFRs in here

32:14and the secrets over there and this do this.

32:17And oh, the SDK is not configured correctly

32:19because it's the YAML and not the NFR and blah,

32:22like all of the glue code, right?

32:23It's the worst part of programming.

32:25But now with agents, when people talk about Vibe coding,

32:28that's kind of the shit that's left for you to do

32:31because the agent gets to have the fun and write the code.

32:35And now you have to do the fucking secrets management

32:37and the deployment and this stuff.

32:38Yeah.

32:39And a lot of the time the agent can't do that

32:40because Google's OAuth flow, a block, bots.

32:43Yeah.

32:44Yeah.

32:44It's like, oh, well, the agent can paint on the canvas.

32:47Now, all you have to do is go and buy the fucking paint.

32:50But what I want is to somebody bring the paint and the brushes, and then I can tell the agent, go and do this.

32:57You know?

32:58Yeah.

32:58And that's what we have when we are building and using AMP.

33:02Yes.

33:02And a lot of the joy, just the sheer wonder and awe that we get to experience is because we get to build on this foundation.

33:10And also, we get so much feedback.

33:12We have tons of users who give us ideas.

33:15and why don't you share the realization we had?

33:20Yeah, the realization is that there's two amps

33:22and one amp is the one that our customers use

33:27and by all accounts, it's a great product.

33:31They love it.

33:31They have fun with it.

33:33But then the amp we use internally

33:35is the amp where I can wake up on a Tuesday

33:39and think I want my orbs to be colored in the sidebar

33:43and send a threat, send off an agent.

33:46And then I can say, yeah, let's put this behind a feature flag and ship it.

33:50And then I post an outslash.

33:51What do you guys think about this?

33:52Then I tweet it out and then I get 10 people saying I want this.

33:55And then I can enable it.

33:57It's from idea to deployed version, it's really, it's short.

34:03Like the biggest blocker right now, I would say is our deployment pipeline.

34:06That's still because we have our scale is now large.

34:11so it takes a while to roll that stuff out safely to a bunch of these um processes running usually

34:18like 10 minutes though well that's yeah there's some other stuff for other stuff like example

34:25yesterday like again like shout out to ian he was like hey can you build this for me i sent off the

34:29prompt it did it and then it was like my 8 30 i had to put the kids to bed and whatnot and i was

34:36Like, I really want to get, like, I really want to tell them it's done.

34:41But then I was like, okay, it's 8.30.

34:43Got to put the kids to bed.

34:45I'm going to bed then.

34:46Like, I don't want to roll this, you know, like that was the biggest thing, but it was

34:50done, done, you know, like it was not done, done, but it was done.

34:53So yeah, that's, that's a realization that AMP for us is what we've previously called

34:58jellyware, right?

34:59It's, it's high quality, it's high tastes.

35:03And I'm quoting here, right?

35:05I'm not dragging.

35:07And yet we don't develop in a traditional sense of how other people develop software.

35:12Like there's a lot of freedom and a lot of trust in this team on how and when people ship stuff.

35:18And it's not vibe coded.

35:19Like it's not slop.

35:22But it's like when we use AMP, we are like homeowners.

35:26And with homeowners, you can make your house your own.

35:30You can put whatever on the walls.

35:31You can paint it however you want.

35:33You got to do all the maintenance.

35:34And if you see something that's wrong, no one else is going to fix it.

35:37You got to go fix it yourself.

35:38And that's how we feel.

35:40And there's a certain joy in that responsibility.

35:44And then the other way of using software is where you're just an end user and you cannot go change it when you're renting an apartment.

35:52And it nice that you can call the landlord and they can go fix the washer and dryer and you don have to deal with that But hey there a lot of stuff you like to change about it and you can and there pros and cons to both but all of a sudden it become

36:06so much easier to go and change the house and so much cheaper and now you know a lot more people

36:12understandably want to be homeowners and so we're thinking about for us how we can get that amp way

36:19to more people and how we can give that to you in AMP and also how we can help you get that in more

36:27of the software that you're using. And it does require change in how companies work. Like if we,

36:36instead of building AMP the way that we do, which is small team, a ton of ownership,

36:41the ability to fix things, to change things, not a lot of process. If instead we had a very slow

36:48issue product management roadmap process and then we had a very slow code review and deployment and

36:54approval process well then it wouldn't matter that we have the ability to change stuff about

37:00amp because we just wouldn't be able to we'd be bottlenecked on either side so yeah we want to

37:05you know as one thing export the amp way to more teams share more about what that means and why

37:12the joy of being able to push to main and why that can lead to higher quality and what you need to do

37:18differently on your team to make it so you can do that in a high quality way yeah that's what lies

37:25beyond the orbs right like what's the next steps like that's yeah like orbs are just the next it's

37:31the next stepping stone right to to fully leaning into this and it's funny by the way someone who

37:39has terrible process on either side for them if it takes you a week to ship something and three

37:45weeks to figure out what to ship, then an EC2 cloud with a bunch of shell scripts is going to

37:51be just as good as AMP. Neither of them are going to be that good. AMP really shines when you can

37:56move so fast. That's when the multiplayer stuff, the portals, the orbs, the ability to do so much

38:01at once, the ability to have it monitor into production because it pushed and it is rolling

38:06it out. All of that stuff, we built AMP for that. So also us exporting the AMP way will help more

38:12people see why orbs are different it's only because you just brought it up i just realized

38:19like how seldom i now think about these companies with traditional processes you know where we get

38:25into contact with them when they said like can you make orbs work for us with in this workflow

38:32in this environment and man like i guess like every software company should throw away

38:37even as a thought experiment

38:40how they've done things

38:41and then restart from scratch

38:43and think like,

38:44okay, if we take the trajectories

38:46that we just talked about,

38:48intelligence will,

38:49prices of intelligence will go down,

38:50speed of tokens will go up,

38:52the intelligence itself will get smarter.

38:55If we take these trajectories

38:56and take them as a given,

38:58how would we develop software?

39:00And I think a lot,

39:02the outcomes would look completely different.

39:04That's the whole,

39:04that's the summary of this conversation, right?

39:06But the funny thing is, the interesting thing to me is when we see companies, we mentioned Linear, two others come to mind like Raycast, Shopify, for example.

39:19They're leaning heavily into – they were founded pre-2025, even maybe 2020.

39:26I don't know.

39:27But like pre-agentic companies and they're leaning heavily into AI.

39:32and so you see them how reshaping the company to this but the weird thing here is that

39:41you know it's selection bias because we don't ever hear about the companies that don't reinvent

39:46themselves like they just die a quiet death like i'm not calling out a name but i found a company

39:52i posted it now flag maybe i could say it's about email but it was like the website looked like

40:0020 you know 2004 like 2008 maybe and it's like that's not like who are you selecting you're

40:10selecting for people that use their computer like it's 2008 and you it's not like they will go out

40:18with a big bang they will go out with a whimper because the world has moved on and it's just not

40:24going to be interesting so i don't know where i'm going with this i'm just saying like you need to

40:28Well, it's easier said than done.

40:30Just pay everyone in the world, go rethink everything.

40:34But that's why as we are thinking about, given that we have AMP, how can we have the biggest impact?

40:42And to be clear, it's not about us suddenly making AMP bend over backwards to accommodate these old workflows.

40:49No, we want to evangelize this new way of working and also take the fact that a lot of the software that these companies use is you can now replace it with something that has 10% of the features, no setting screen, is maybe 1% of the cost, and is infinitely more customizable.

41:08and you don't have to be a company that is ready to be agent pilled.

41:13You don't have to be someone who's using the agents directly

41:16in all the ways that we might want you to

41:18in order to get a ton of value from that

41:20because you can go and say,

41:22I used to have this CRM where, you know,

41:26one really smart person on my team had to manage it

41:28basically, you know, all the time.

41:30It was super clunky.

41:31It was super slow.

41:32And now I can just say,

41:35hey, I want you to connect to Meta and pull in the data from here for our ad campaign.

41:41And that is something that anyone can appreciate.

41:45So it's how can we get this impact out more?

41:49And that's why I don't think that on the margin, the hundredth person building their own agent platform product,

41:57I don't think that's going to move the needle.

41:58I think it's time for us at AMP and really software developers to think about how you can actually solve the problem, not build software to build software to solve the problem.

42:08Yeah.

42:10And you just reminded me of something or made it click, which is when, you know, in that triangle of build versus buy versus build and buy, I don't know what the, but there's a thing here where what you just said, right?

42:24Like it's easier said than done rethinking your business constantly.

42:27while the models are getting better every two months, right?

42:31And while the costs are changing

42:33and how good they can do different things is constantly changing.

42:38So this is what we're doing at AMP.

42:41Like if you're using AMP and paying us,

42:44you're literally paying us to constantly rethink

42:46how software is being done.

42:48And this is what I do all day long, right?

42:50And that's in the product.

42:52The way we think about how software can be built nowadays,

42:55that's the product.

42:56That's an expression.

42:57And it will look different in two months, right?

43:00So that's another thing.

43:02So in times of these technological changes, it's not just like a static build versus buy because the landscape is completely changing, right?

43:11It's like, you know, do I make my own rain jacket or do I buy a rain jacket?

43:16No, no, no.

43:17It's like you're going in the jungle and there's like a guy there and he's going to give you the clothes for the day because he can read the weather there.

43:25It doesn't matter whether you buy or build your own rain jacket.

43:29The weather is going to change and you want the guy in there who knows, like, can read the clouds and whatnot, right?

43:35Not to oversell it.

43:36But I think that's – I don't know what that analogy is, but, you know, like AI land is a jungle.

43:42I think if you import that to software, in the past, software was boring.

43:47It had to be mediocre.

43:48Because of that principle of the software business, you build it in year one and then you milk it for the next 10 years by –

43:54because it's zero marginal cost copying you know you want it to appeal to the broadest market

43:59possible and now that's changing like why do i use ghosty because i trust mitchell in yeah what he

44:07thinks a terminal emulator should be i want to see what is the issue tracker not made by a company

44:16that has to water it down to appeal to a lot of people but what is the issue tracker made by

44:20someone that I really respect that has crazy ideas who's going to show me not just oh it works

44:27but how I should be doing this in the future or how we do video calls how we plan our team events

44:35how we do travel all these and so in the past if you had a great idea here you would have to go and

44:41go hire a bunch of people spend all your time doing that not on the product you'd go raise VC

44:46And in the end, the incentives would not be to build the most opinionated, correct product.

44:52And there would only be like a few competitors now.

44:54There can be so many more competitors.

44:56And you could even have, I want to use so-and-so famous developers or famous product person's product because their name is what gives me trust in it being different.

45:07And they can make so much more money, and it can be cheaper because there's so much less overhead that they need to go and build that and distribute that and sell that.

45:16they do not need to make a setting screen.

45:19They do not need to make it multi-tenant.

45:20They do not need to do all of this other stuff

45:23that they used to have to do.

45:24And that is really exciting.

45:27So in the same way that we want AMP

45:28to be a representation of how we think

45:31you should be building.

45:32And if you trust us, then use AMP

45:34and we think you'll have a great time.

45:35I want that for all the software that I use.

45:38Yes.

45:40You're paying somebody else

45:41to think about the problem constantly.

45:43That's what you want.

45:44I want Mitchell out there

45:45thinking about terminals constantly. I want whoever the person is at Google Meet who thought

45:51of when I switched tabs to show me the little pop-up, you know, where the Google Meet meeting

45:55goes on. Great job. Like, I want you to think about this. I don't want to think about it,

45:58which goes back to your point of like, you want to do the thing and not build the thing to do the

46:03thing. And yes, intelligence is cheap and tokens are cheap and thus labor, like intelligence labor

46:10is cheap, but like

46:13the thinking that a human does about

46:15a problem space is still different, right?

46:17Because it's continuous and they

46:18have different options, they weigh different options, they're human.

46:21So that's kind

46:23of what's

46:24worth it, right? That's what you're paying for.

46:27I want somebody, I want like

46:28a Rasmus Anderson design text editor

46:31or whatever, because I don't want to think through the edge cases

46:33or Zed or whatever. Like I can whip up

46:35my own, but for certain

46:37things I want other people to think about it.

46:38yeah oh yeah i think we should we should any i'll end it with a short anecdote because you brought

46:45up ghosty and the terminal so we're all we all killed our local dev i don't know barely use it

46:51anymore right and um i posted an outside i said like in the last six weeks i forgot all of the

46:57jujitsu or jj shortcuts that i set up the aliases i was completely lost and then tim culverhouse

47:04replied and Tim Coverhouse

47:06for those who don't know

47:07he built

47:09his own terminal emulator

47:12he built his own I think Wayland

47:14compositor window manager he built

47:16his own email client he built

47:17like everything he uses he built himself

47:20he owns shell he has his

47:22own shell running in his own terminal

47:24and he replied saying I have

47:26my I have the best terminal and the best shell

47:28in the world and I don't use them

47:29and it's and a sad smiley

47:32because it's fully in Orbs now.

47:34So what does that say about the value of software?

47:39Yeah, and everything is changing faster.

47:42So if you were someone who has thought X

47:44and then you found X to not be true,

47:46I mean, we have, and it's very humbling,

47:48then you just get your mind to a place

47:50where you permanently are more open

47:52of crazy changes out there.

47:56Yeah, wild times.

47:58All right, we'll see you next time.

0:00Quinn SlackJust building software to do X is only valuable if you can get people to, to pay you, and I don't-- I think that that's gonna be a lot harder given that people can just make the software themselves. So developers need to think of themselves as doing X, not building software to do X.

0:15Thorsten BallYeah.

0:15Quinn SlackAnd I think that's gonna be a rude shift, because if anything, software developers on average have shifted to be more general-

0:21Thorsten BallYeah

0:21Quinn Slack... more horizontal in what they're building.

0:23Thorsten BallThe richest companies on Earth are pouring all of their cash into this, and we're not even seeing the results of this yet. Like, this is still being built. If you now think, "Oh, I, I have to be, uh, uh, stingy with my monthly subscription and I cannot run out of tokens," I think in a year or two you will find that the

0:44Thorsten Balltokens are everywhere. It's like electricity. It's like you're not measuring the watts coming out of your wall. And I think it's the same thing with tokens. There will be so many tokens everywhere and so fast that people won't recognize computing anymore.

1:06Thorsten BallAnd we're back. Another episode of Raising an Agent. This time we're back home. Uh, one of us in, in San Francisco in Mill Valley, the other one, um, in this really small town in Germany near Frankfurt Airport. Um, and yeah, we're excited to talk again, right? Like, Munich sparked a lot of ideas. Munich got us, get the,

1:25Thorsten Ballthe stuff flowing in our heads, the jelly flowing.

1:29Quinn SlackYeah. We've been shipping a lot, and it's been a few weeks since Munich, and we had said then that the way that we build had changed more in the prior like four or six weeks than it had changed all of last year. And think we're still feeling that.

1:43Thorsten BallYeah.

1:43Quinn SlackAnd we're starting to think about looking even further ahead and what is the world going to be like and how can we take what we're building, what we're seeing and get more people building in the AMP way. And also then thinking about what is that going to do to software and what that means for software, right?

2:03Thorsten BallYeah.

2:04Quinn SlackSo-

2:04Thorsten BallYeah

2:04Quinn Slack... Thurston,

2:06Quinn Slackw- predictions. What do you think is gonna be different six, nine months from now?

2:13Thorsten BallI think all of the trajectories will continue. Like, I don't think there's gonna be this rug pull that most people say there might be or there will be. So I think the price of intelligence will drop. And people are saying, "Well, the models are more expensive." That's not... Well, the newest ones, right? But like the price

2:31Thorsten Ballof the old models has dropped considerably. And I think in, say, half a year, the price of GPT-5, 6 levels of intelligence will be dirt cheap compared to right now. And you know, Fable, incredibly expensive. I think intelligence at that level will be cheap. I also think if you look at the trajectory of tokens per second,

2:54Thorsten Ballright? Uh, OpenAI is rolling out their super fast mode. Um, uh, I ha- don't have access. I wish I had access. It's, it's... What is it? 750 tokens per second, which is mind-blowing.

3:06Quinn SlackAnd if we had access, we would not be able to say, so.

3:09Thorsten BallI, yeah, if I had access, I wouldn't be here talking, to be honest. Like I, I wouldn't send tokens. It-- That's mind-blowing. So I think this is... I'm not sure if this is correct actually, but I think it's close to like the build-out we're seeing with AI, the data centers, the GPUs. It's one of the biggest infrastructure

3:29Thorsten Ballbuild-outs in the history of humanity, you know? Like this, the richest companies on Earth are pouring all of their cash into this. And we only just-- We're not even seeing the results of this yet. Like, this is still being built. So if you now think, "Oh, I, I have to be, uh, uh, stingy with my monthly subscription and

3:51Thorsten BallI cannot run out of tokens, and I count the tokens, and you know, 50k tokens is, uh, something that I measure," I think in a year or two or maybe three you will find that the tokens are everywhere. It's like electricity. It's like you're not measuring the watts coming out of your wall. And you won't measure... You know,

4:09Thorsten Ballyou measure them in fucking megawatts and kilowatts, and you don't care about like 50 watts, right? And I think it's the same thing with tokens. There will be so many tokens everywhere and so fast that people won't recognize computing anymore. I think based on these projections, a lot of things that we think about how software

4:33Thorsten Balldevelopment is will change. Well, do you disagree with any of these predictions or do you want to add some?

4:41Quinn SlackI agree. Intelligence is getting cheaper and cheaper.

4:46Thorsten BallYeah.

4:46Quinn SlackAnd anyone who thinks there's gonna be a rug pull, the quote-unquote "subsidies" will end, there's no evidence that subscriptions, which is what people usually mean, are actually unprofitable. And it's hard to actually tease it apart in the same way that it's not like United Airlines is going to stop having economy seating

5:04Quinn Slackbecause that's been a subsidy and they're gonna pull the subsidy away. No, you know, once there's a big upfront expense, it makes sense to spread that in a price-discriminated way across the largest pop- possible population of buyers.

5:15Thorsten BallYeah.

5:16Quinn SlackAnd also totally agree that if you are in a role where you're having to be really stingy with the intelligence that you can use, then that's a signal. In much the same way if you're someone who's getting paid one-twentieth of what a San Francisco engineer makes, then find a way to get to San Francisco or closer to San Francisco.

5:36Quinn SlackIt's a signal that what you're doing is not benefiting enough from tokens and you're not creating enough, uh, value. You're not in that good position. But if you're thinking that, then you're probably the kind of person who can create a ton of value if you get into a job or other setting-

5:51Thorsten BallYeah

5:51Quinn Slack... where the tokens can flow. And at the same time, there's a lot of noise going on right now. First, it feels like everybody is ... trying to build the same thing. Now, you know, Thurston, you and I have been building dev tools for a long time.

6:09Thorsten BallYeah.

6:09Quinn SlackAnd for the longest time everyone said, "Oh, you can't make any money in it."

6:15Thorsten BallYeah.

6:15Quinn SlackAnd now people have completely changed to the point where it feels like everybody is building dev tools, and that also is probably not correct.

6:26Thorsten BallYeah.

6:26Quinn SlackAnd ultimately, software, it has to solve the problem. You know, just, just building software to do X is only valuable if you can get people to, to pay you, and I don't-- I think that that's gonna be a lot harder given that people can just make the software themselves. So I think more and more, we've said this before, but

6:46Quinn Slackdevelopers need to think of themselves as doing X, not building software to do X.

6:50Thorsten BallYeah.

6:50Quinn SlackAnd I think that's gonna be a root shift, 'cause if anything, software developers on average have shifted to be more general-

6:56Thorsten BallYeah

6:56Quinn Slack... more horizontal in what they're building.

6:58Thorsten BallI do have a little pet theory that I wanna offer here, and it's, it's a pet theory. Say the inflection point was November 25, right? Which this is my other pet theory. It wasn't for us. I don't think November 25 was that big. I don't think there was a bump in the curve. I think it was a sm- curve and we could-

7:16Quinn SlackExcept for Gemini, remember?

7:18Thorsten BallThat, that was... Well, let's not talk about it. But I think, um,

7:23Thorsten BallI think it was a smooth curve, but, um, there's a lot of engineers who've kind of woken up to agents now, and now we've seen like in the last few months, um, people... And I've said this before on this podcast, but it's, it's even more than this, where people are now like they're hacking on their own agents and like they'll

7:43Thorsten Ballplug in. And it's-- I thought this would go away, but it's still a thing that's going on, where people build their own tooling around this stuff and blah, blah, blah, blah, blah. And my pet theory is engineers have this thing now where

7:58Thorsten Ballthey, they go, "It can't be this easy," right? Like, you know like how we always try to fix our tools and build like some elaborate workflows and automate stuff? Like I mean, not me. I'm the guy who copy and pastes, you know, I'm, I'm not doing this. But like if you send a prompt to Fable, I did this yesterday, I used Fable.

8:14Thorsten BallI'm kinda on a

8:17Thorsten Balltwo-day Fable binge. I don't know. Maybe it's because I wanna spend money after we can put ChatGPT subs into Amp, and now I'm like, well, you know, gotta not, you know, gotta use Fable. So I send a prompt yesterday. Ian Landsman, shout out to Ian. He sent a tweet. He's like, "I wanna be able to set the default branch in

8:34Thorsten BallAmp for Orbs." So I send a prompt, I tweeted the prompt. It was like a text message I sent, I would send to somebody. And Fable went and did it, and it just added a feature. It didn't add complexity. It kinda wired that new field, the default branch, through like our different services. I had it watch deployment. I had

8:53Thorsten Ballit make sure it rolled out. I tested it and it works. I do think, and here's the pet theory, that engineers then go and look at this and go, "Oh, it can't be this easy," right? How about I add a bunch of stuff? I add like some tools and like some shortcuts and like a review thing and this process and that process and blah,

9:12Thorsten Ballblah, blah. Because it doesn't appeal to the engineers of the last, say, 20 years that you just say something in plain English and it goes and does it. You know? Like I think there's like something like this going on where you wanna kinda add like some value and that's, that's how you build like these other engineering,

9:30Thorsten Ballth- this whole elaborate stuff around it, but-

9:32Quinn SlackYeah

9:33Thorsten Ball... most of the time it's not-

9:34Quinn SlackWell, it's like if you're a furniture builder, you would probably feel ashamed if someone comes over to your house and it's all IKEA furniture.

9:40Thorsten BallYeah.

9:40Quinn SlackAnd when you sit down on that IKEA couch, you're gonna hear every little creak and it's gonna annoy-

9:44Thorsten BallYeah, yeah, yeah

9:45Quinn Slack... the hell out of you.

9:46Thorsten BallI mean, we all know the story, right? That some guy posted, I don't know, 20 years ago on the internet. They, they were a 3D designer, and basically they figured out that every time they had like something they made and they showed it to their boss and said like, "Does this look good?" The boss would always say j- and just

10:03Thorsten Ballto, the boss would feel they have to add value, so they would ask like, "Change this," even though it was completely nonsensical. So the designer added a little duck to everything, and then the boss would always go, "That looks great, but get rid of the duck." That duck, uh, and then they would feel happy about it, they

10:21Thorsten Balladded something. So I do think there's, there might be something going on here where engineers feel like it can't be as easy as just saying something to the model and build software, right? And so they have to add something.

10:33Quinn SlackYeah.

10:33Thorsten BallBut yeah.

10:34Quinn SlackSo with an, an agent though, you wrote the how to build an agent that turned on a lot of people to how simple the idea of an agent loop with tools-

10:43Thorsten BallYeah

10:43Quinn Slack... could be, and that was pretty cool. That opened people's eyes.

10:49Thorsten BallYeah.

10:49Quinn SlackThat's a new primitive. What about all these people who say, "Oh, Orbs, how is that different from just EC2 and you set up a VPN and you have some orchestration?" You know.

11:01Thorsten BallYeah.

11:01Quinn SlackLike what are, what are they gonna run into that they can't build themselves?

11:07Thorsten BallThere's multiple things going on. I think, well, first of all, we have to admit that building stuff is pretty easy now and getting started with stuff, right? And we had like the first version of what's now called Orbs

11:21Thorsten Ballin two weeks, a week or something. And then we ran into the long tail with that infrastructure provider we used and of issues and, you know, whole slog. And then I, I think I rewrote it on the new platform over the course of a weekend. And this was not me sitting here for 12 hours. This was me starting prompts and whatnot.

11:42Thorsten BallAnd so admittedly, it is easy to build a lot of stuff. But when somebody says, "Oh, I built Orbs and it's like EC2 and a VPN and, uh, or you know, I have like 15 Tmux sessions running on my home lab and like 18 different cloud codes in it," and you can get far, but it's obviously not the same experience. Like the, the value

12:05Thorsten Ballof Orbs is that you have All of the things in the same package, you know. Like it's not... You don't have to manage Tmux sessions. The beauty of Orbs is you don't have to think about the machine. You just send your prompt, you get the results back, you preview it in a portal, you can share it with your colleagues and whatnot.

12:22Thorsten BallLike you don't have to worry about how the sausage is made. And there's also people who are offended by this where they're like, "I wanna see how the sausage is made," you know. But I don't know. I don't know what to say. Like, uh, it's the same thing maybe. Like are you really concerned about how we spawn up these sandboxes

12:38Thorsten Ballor VMs? Like do you... What do you want to see? You know, like you can specify hooks and scripts now. You can define your own Bash script. We're pretty ex- exposing a bunch of stuff. But I, I think it's underselling the whole experience. Like you can get like 80% of the way there, but what we found is the 20% makes a lot

12:56Thorsten Ballof difference. It's weird how hard it was to get this nailed down because it's not like super hard programming. Also, what does that mean anymore? I'm not going on this sidetrack, but you know. Like if you'd asked me five years ago, super hard program would be like an optimizing pass for an optimizing compiler, you know.

13:16Thorsten BallI'm pretty sure that has kind of flipped now. Like-

13:19Quinn SlackYeah. It's like the design and-

13:21Thorsten BallYeah, the design and-

13:22Quinn Slack... how all these pieces fit together.

13:24Thorsten BallYeah, into a whole and like the product. Who, what are you building? In what sequence? When are you building it? What are the trade-offs? You know, all of this stuff. So when we

13:36Thorsten Ballbuilt this, we f- th- this was really hard. Like how... What are the primitives, for example, and what are the things that, the invariance of the product? One thing that I think we got right is

13:50Thorsten Ballyou start an Orb, you get a single URL, you get a fresh Orb for new threads. That was controversial when I said this, right? Like everybody was like, "No, no, no, I want to spawn multiple agents in an Orb," you know, in the same thing. And now look at us. Like that would be different product where you can spawn multiple

14:05Thorsten Ballagent. Like, you know, like with different and select a specific Orb and whatnot. That would be a different thing. So people can go far and wire their own stuff together, but wiring your own stuff together is not building the same product, and there's a lot of like decisions that are not code level that go into making the

14:25Thorsten Ballfull thing, and those seem to be valuable, right? If you look at the feedback we get where people say it's, it just works or it's the great experience or, uh, somebody tweeted this just an hour ago, like, um, the f- your fr- UI or your like your front end is completely on a different level and whatnot, right? And

14:45Thorsten Ballobviously, like we have great designers and engineers on the team who, who really sweat over the details here, and I don't know. I don't know. I, I don't know how you can replicate this with your homemade VPN EC2-

14:58Quinn SlackYeah

14:58Thorsten Ball... the Tmux thing.

15:00Quinn SlackI, I think there's-

15:00Thorsten BallWhat do you think?

15:00Quinn Slack... two ways to look at it. One is if you're just trying to have fun or just build something for yourself to use and you don't need to make it work at like team scale, you don't need to get other humans involved, you don't need to deal with the complexities that come with other people also wanting to make it work their

15:19Quinn Slackway, then, you know, you could probably build something like a-

15:23Thorsten BallYeah

15:23Quinn Slack... agent platform that's pretty fun.

15:26Thorsten BallYeah.

15:26Quinn SlackBut once you start making it so that an entire team can work on it, that all the different ways they use it work, that you can read other people's threads, you can have the, you know, Puck-like orchestrator agent work across everything in it, that all the concepts fit together-

15:43Thorsten BallYeah

15:43Quinn Slack... that's hard. And the hard part for us has been no single person on our team, you know, our team is 20 people, all of Amp is 20 people, no single person on our team has every single piece of the system and the design in their head. And so much of it is figuring out how we can have all the different pieces that we all

16:03Quinn Slackknow and feel really strongly about work well together.

16:05Thorsten BallYeah.

16:06Quinn SlackAnd that is tough. We all use the product all the time. We have a small team that's highly trusted, so we've optimized Amp for being able to do that. But that's where if it's, you know, you don't really have that need if you're one person building it for yourself. And if you're one person-

16:21Thorsten BallYeah

16:21Quinn Slack... building it for your whole company, you're gonna run into all the same problems that we have. And then-

16:27Thorsten BallYeah

16:27Quinn Slack... people are gonna be saying, "Well, why can't I use Amp?" Or, you know, all the other... Now there's 87 different things that are in this category. And that is probably gonna give you-

16:36Thorsten BallYeah

16:36Quinn Slack... headache. But the more important thing is, like, what is... It's fun. I-- Don't get me wrong. That's part of why we do it. But you're not actually creating value. It's not like any of these agents like Amp, it's not like we're charging you a ton for the joy of using our product.

16:56Thorsten BallYeah.

16:56Quinn SlackWe're making money. We're profitable. But you're not gonna save a lot of money. You're probably gonna actually lose money on net 'cause there's a lot of efficiency stuff we do. So what is the business value you're, you're doing? You can't be making software to do X. You most certainly cannot be making software to make software

17:10Quinn Slackto do X. You gotta be doing-

17:12Thorsten BallYeah

17:12Quinn Slack... X these days.

17:13Thorsten BallYeah. Yeah, but I mean, it points at a g- at a real tension, right? The whole build versus buy thing, right? Where should we buy something and not build it ourselves? And two years ago, you know, besides operational costs and like guarantees and support and whatnot, you had the big thing of, "I don't wanna build this myself,"

17:38Thorsten Ballor, "I can't build it myself." But now in the age of I take a screenshot and I say, "You know, build this for me," that equation has changed. Like that scale has been tipped, and I think the interesting bit is that the models are now weight on the scale, on the build versus buy scale, and, and the weight of the models changes

17:59Thorsten Ballconstantly. Like it's, it's, you know, like you constantly have to reevaluate. And I mean, admittedly, I built a bug tracker into Amp last week, right? And it was one-shotted and it was deeply integrated into our system. So- That might have been even easier than spending the time to setting up Linear or something because

18:19Thorsten Ballnow the value is in deeply integrating stuff, which is why we're even talking here, right? Like this whole thing is, is incredibly fascinating to me that if you look at the history of software, and we've said this 1,000 times, software was hard and expensive to write, slow. Basically you want the investment you make when

18:38Thorsten Ballwriting software to pay off or have a return, so you generalize the software that you write so you can sell it or reuse it multiple times in different use cases, right? You don't build one-off software. That's not... Doesn't pay or it hasn't paid. But-

18:55Quinn SlackYeah. The beauty of a software business is you build it and then for the next 10 years-

19:00Thorsten BallYeah

19:00Quinn Slack... your customers will stay on it and you can sell it and infinite copies can be made at zero marginal cost and you can make money on each of those.

19:08Thorsten BallYes.

19:08Quinn SlackSo you recoup your fixed expense.

19:10Thorsten BallYeah. Even, even if you're an agency and you build customer specific things, most agencies have then developed some kind of framework or like some, I guess you can call it a stock or whatever, like something that they reuse across projects, right? Which means they can save time, like a common admin interface or whatever.

19:27Thorsten BallOr, um, think of like WordPress, right? Which is, I mean it's been 10 years since I checked, but back then it was most configurable pieces of software, like the admin interface had a lot of menus and now it's like, well you could take a screenshot or four of different websites and have like one shot a thing, like your hair

19:49Thorsten Ballsalon website or like your whatever restaurant website and then every time you wanna make a change you can ask the agent again to do it, right? So this has changed and, and, and like what I find fascinating is that a lot of practices and methodologies of software, we still do them or, or they're kind of crumbling, but um,

20:11Thorsten Balla lot of stuff doesn't hold anymore, right? Where-

20:14Quinn SlackYeah

20:14Thorsten Ball... we talked about this-

20:15Quinn SlackSo before you even get onto that, I think you mentioned this but I just wanna put it to you again. Basically what we are saying is we see a lot of people say, "Oh, AMP or other agents, I could just replace that with EC2 and some shell scripts-

20:28Thorsten BallYeah

20:29Quinn Slack... and get the same thing." And, and we understand that but we don't think that that's the right thing for them to do. But at the same time you're saying, "I was able to code the parts of Linear, an issue tracker that I need from scratch."

20:41Thorsten BallYeah.

20:41Quinn Slack"I, you know, quote unquote replaced Linear with a few CRUD forms." So which is correct? How could we say that in one case, you know, AMP is worth it, but in the other case no Linear is not worth it? What, what makes it different?

20:56Thorsten BallThat's a good question. I think what's really different about...

21:03Thorsten BallLet's start on this end. I would say the Linear or bug report case is so different because the workflow around how you do bugs is different in each company and it's like the, it's, it's like bug tracking or issue reporting or knowledge management, like it's different in every company. That's why Jira has 15,000 options,

21:27Thorsten Ballright? And why every Jira is different. It's not because people don't like the button but it's because they, they work in a different way. So I think like-

21:35Quinn SlackCouldn't you say the way that people use agents and do software development is different in every company and therefore every company should have their own-

21:40Thorsten BallI mean I guess you could. I'm starting to, I'm, I'm thinking about it. But I would argue that the, uh, it's like the thing I said at the start, right? Like you have like, you need to specify what you want and the model does the thing and there's not a lot of this... It's handed over to the model is I guess what I'm saying.

21:57Thorsten BallThere's not a lot of touch points anymore with the humans. Maybe that's the difference in that,

22:03Thorsten Ballyou know, like previously like your dev environment had like a lot of touch points where the human were ch- the humans were changing things. But now man, like in orgs like they just whip up software and tear it down and I don't care and I don't know. What do you think? Like this is, is interesting. What do you think?

22:19Quinn SlackI think this is the big question that everyone has about software-

22:22Thorsten BallYeah

22:22Quinn Slack... is what software is below the waterline where-

22:26Thorsten BallYeah

22:26Quinn Slack... you don't wanna go dive and mess with it and what software's above the wa- waterline where you wanna, you know, go, go build. Um-

22:33Thorsten BallYeah

22:34Quinn Slack... if you just look at people's behavior I think you've seen a lot of people shift from wanting to spend their time in issue land and project-

22:46Thorsten BallYeah

22:46Quinn Slack... land and typing that in and now they want to spend their time in agent land and you cannot argue with people's revealed preferences, their actual actions. And so there's something about, well why do we think AMP or generally the agent is different from the issue tracker? It's because we have found it to be incredibly

23:04Quinn Slackproductive to work there, far more than issues ever would. And so there's, there's more value to having that be really good and solid.

23:14Thorsten BallOh. I see what you're saying. So you... Okay, I see what you're saying. So the value, the business value of the software has changed and you would say that building your own for low value things is, that's, that's, you know, like the-

23:29Quinn SlackWe don't spend that much time in it ourselves so-

23:32Thorsten BallYeah

23:33Quinn Slack... there's not that much complexity that we need to replicate. So in the past maybe we would've-

23:37Thorsten BallYeah

23:39Quinn Slack... used 60% of the features of Linear and Linear-

23:43Thorsten BallYeah

23:43Quinn Slack... is an awesome, awesome product. You know, we, we've used it and it is-

23:47Thorsten BallYeah

23:47Quinn Slack... the best of its category but now-

23:50Thorsten BallYeah

23:50Quinn Slack... we might only be using 10%. And I think there's a lot of people that are using some of Linear's agent features and they like that, but we would only be using 10% of it.

23:59Thorsten BallYeah.

23:59Quinn SlackSo I get, I think there's also an element where Linear was built for a workflow that is undergoing a lot of change and if you ha... I think there's also something fundamental about software businesses that were started before 2025 and have not completely reinvented themselves and ... burn the bridges.

24:20Thorsten BallYeah.

24:20Quinn SlackI think there's also a lot of valid skepticism that, that those workflows will not be, be there.

24:27Thorsten BallYeah.

24:27Quinn SlackI think there's also something about if, like, an agent is at the core, that then that just makes the software so much more pliable and moldable. Like in the way with Amp, if you wanna do something-

24:36Thorsten BallYeah

24:36Quinn Slack... most of the time you can just ask Puck to do it. If you want an integration, you can just ask, you know, Puck or Amp in a thread to go set up a web hook.

24:45Thorsten BallYeah.

24:46Quinn SlackAnd yeah. I don't know.

24:46Thorsten BallThe, the-

24:46Quinn SlackBut, you know, this, this is what we're thinking about right now.

24:48Thorsten BallYeah. Another angle is the whole f- I don't really like the term, but, like, the whole system of record type of thing, meaning, like, some software is more important than others. I think that's kinda what people wanna say when they use the term, meaning with orbs, you, you're spending money on VMs, so as soon as you're

25:06Thorsten Ballspending money, like significant amounts of money, that changes this whole equation, right? And you have the tokens flowing through the system. You wanna keep track of this. I do think, like, the whole process of that touches the code is important, so, uh, you know, like there's something systemy of record-ishy about, like,

25:26Thorsten Ballthe orbs, whereas with the tickets thing that I built, it's not, right? Like, for us, that's not a cost. Like, it's, it's, it's a CRUD app in our admin CRUD app. It's basically free, and that changes it for me. And, like, the whole orb stuff, that's our main application. That's the main thing that's, that's load-bearing,

25:45Thorsten Balluh, to use the, the, the stolen term the agent stole from us. But I don't know, maybe that's this. But I do think, like, what you said is very interesting with you have the agent at the core, and you can change it, and the, there's obviously now this tension of Linear was built in a time where you had to, for different

26:06Thorsten Ballcustomers, they, they have different workflows, so you had to have the option to custom labels or, like, custom workflows or, like, custom boards and whatnot. And a- admittedly, Linear was, I think three years ago, they were pretty minimal, right? And they said, like, "We only have this." And then people said, like, the

26:23Thorsten Ballfo- look, I don't know, that's not true. But, like, couple years back, they started adding more features, and it became more complex, complex. And obviously, if you're, if you're SaaS software, you have to be this complex because you have multiple customers, and they give you different amounts of money, and they have different

26:40Thorsten Ballusage and different workflows. So you, your software has to be generic for different things, right? But with agents now, if you build your own, it doesn't have to be generic at all, and that kinda changes the whole cost thing of the software again. Like, it makes it less complex and whatnot. But obviously, like, what I'm

27:01Thorsten Balltalking about here is, like, this triangle of, well, you know, you can buy the SaaS thing, and it's complex in the back and a lot of work to build because you have to generalize it, or you build your own, and then it's kinda maybe cheap to build, but it's customized to your workflow. But then on the other end of the triangle,

27:19Thorsten Ballthere's this thing that, you know, with orbs and whatnot, where it's, it's load-bearing. It costs money. You do wanna modify it, but

27:29Thorsten Ballyeah, I don't know. Like, I think everybody's trying to figure out where-

27:31Quinn SlackYeah

27:31Thorsten Ball... what goes where on this triangle.

27:34Quinn SlackYeah. And with Amp, we, we kill features that we no longer think are good, so there's not cruft in Amp that we're keeping-

27:42Thorsten BallYeah

27:42Quinn Slack... around because some customer nine months ago paid us for a year, and they'd be pissed if we take it out. No, I mean-

27:48Thorsten BallYeah

27:48Quinn Slack... everyone knows with Amp that there's no cruft in it. And by the way, Linear is, of basically any software company, best positioned here, and they've done way better than pretty much any other company I know in, in staying relevant. So in a sense, we're picking on them because, you know, they're the strongest. They're

28:05Quinn Slackan interesting case. They're a good test case here. Um, and I think if, as we look to the future of what software will be, certainly for the stuff that we use at Amp, if there's something that our agent is the primary user of or where we've ever been annoyed at the UI or their salespeople wanting us to get on the phone

28:28Quinn Slackwith them or anything about that, that's danger territory, and I think a lot of-

28:33Thorsten BallYeah

28:33Quinn Slack... companies are, are seeing that. Like, if you are not, if you're using a tiny fraction of the features of a software package, it is so tempting to just go and build it yourself. So-

28:45Thorsten BallYeah

28:45Quinn Slack... think your dashboard application, and in that case, a lot of the complexity in that is, like, their own programming language on top, and you're not even interacting with that. You get zero value from that anymore. It also is kind of annoying to know that you're, maybe this is not the right way to think about it, but

29:00Quinn Slackyou're paying for a massive feature that suddenly you no longer use.

29:04Thorsten BallYeah.

29:04Quinn SlackAnd it's really tempting, and I do think that people will be doing that more and more, but there's this view, a lot of people talk about, like, oh, you know, vibe-coded Salesforce from scratch.

29:15Thorsten BallYeah.

29:15Quinn SlackAll of that. That, people were talking about that last year. I think that's dumb because it's not Salesforce that you would vibe code from scratch. It's the 2% feature set of-

29:26Thorsten BallYes

29:26Quinn Slack... Salesforce that you need.

29:27Thorsten BallYou're not gonna, you're not gonna vibe code Excel, but you're gonna vibe code that one spreadsheet that gets sent around and that kinda s- I don't, I wanted to use the word load-bearing, and that's load-bearing for your one workflow. Like, that's what you're gonna vibe code-

29:43Quinn SlackYeah

29:43Thorsten Ball... or build it yourself. Yeah.

29:45Quinn SlackYeah. Or there could be a core where someone, maybe at your company, maybe someone externally, has built the really good primitives that would let you then vibe code. Because vibe coding, if that means-

29:58Thorsten BallYeah

29:58Quinn Slack... starting from scratch every single time, even as models are getting better and better, you want consistency. And take, like, Granola, the recording app for Mac I, in doing dictation, uh, on Amp, I've run into a lot of macOS APIs around audio, and there's a lot of foot guns. There's just a lot of complexity. The models

30:17Quinn Slackstill do not get it right regularly. So you want someone who's figured out all of that, but then you could configure your Jellyware granola to route it here and to understand these dictation things, to not send all your data to the other company, all these things. So I think it's, you wanna do it on top of a, a core that's

30:35Quinn Slackgood.

30:36Thorsten BallYeah.

30:36Quinn SlackSo it's not that there's no role for software packages that are made by someone else.

30:41Thorsten BallYeah.

30:41Quinn SlackIt's that that is completely changing that role.

30:45Thorsten BallYes. I, I 100% agree. Like, even with agents, I do not want to think about authentication. I do not. I just don't wanna do it. It's like what I want is, I have an app idea and authentication should just be there. Like, somebody else should think about it, and that's not the agent. That's not a thing that the agent can do.

31:06Thorsten BallYou need management thinking. You need thinking of different options and weighing the trade-offs and whatnot. I don't wanna think about payments. I don't want to think about... When I'm building an application, I don't wanna think about the storage layer, you know? Like, the database or whatever. So-

31:21Quinn SlackYeah. And you're talking about when you're building on top of a core. You want someone to have thought about that-

31:26Thorsten Ball100%

31:26Quinn Slack... so that you don't have to think about it.

31:27Thorsten BallYeah, yeah, yeah. Exactly. And man, like, people say, like, vibe coding is fun. No, no, no. The, the most fun with programming, and we'll, we can say this is what we do with Amp, the most fun with agents and programming is when there f- there is a foundation that's solid and that's working, and you just can let the agent

31:45Thorsten Ballgo loose on it and, and change the UI or different things or add new features or remove features and whatnot without having to bump into those foundations. Like, if you have an app where authentication is set up and the secrets are set up and, like, your production deployment is set up and you just have to tell the agent,

32:03Thorsten Ball"Deploy the new version," then it's fun. But, like, the, it's always been the case that the worst part of programming is this whole wiring that shit together and putting the env vars in here and the secrets over there and this do this, and, ugh, the SDK is not configured correctly because it's the YAML and not the env var

32:21Thorsten Balland blah. Like, all of the glue code, right? It's the worst part of programming. But now, with agents, when people talk about vibe coding, that's kind of the shit that's left for you to do because the agent gets to do all of the f- uh, gets to have the fun and write the code, and now you have to do the fucking secrets management

32:37Thorsten Balland the deployment and this stuff.

32:38Quinn SlackYeah. And a lot of the time the agent can't do that because Google's OAuth flow will block bots and-

32:43Thorsten BallYeah. Yeah

32:44Quinn Slack... all that.

32:44Thorsten BallIt's like, oh, well, the, the agent can paint on the canvas now. All you have to do is go and buy the fucking paint. But what I want is to somebody bring the paint and the brushes, and then I can tell the agent, "Go and do this," you know?

32:58Quinn SlackYeah.

32:58Thorsten BallAnd-

32:58Quinn SlackAnd that's what we have when we are building and using Amp.

33:02Thorsten BallYes. Yeah.

33:02Quinn SlackAnd a lot of the joy, just the sheer wonder and awe that we get to experience is because we get to build on this foundation, and also we get so much feedback. We have tons of users who-

33:14Thorsten BallYeah

33:14Quinn Slack... give us ideas. And we... Why don't you share the realization we had?

33:20Thorsten BallYeah, the realization is that there's two Amps, and w- one Amp is the one that our customers use, and by all accounts, it's a great product. They love it. They ha- they have fun with it. But then the Amp we use internally is the Amp where I can wake up on a Tuesday and think, "I want my orbs to be colored in the sidebar,"

33:43Thorsten Balland send a thread, s- send off an agent, and then I can say, "Yeah, let's put this behind a feature flag and ship it," and then I post it in our Slack channel, "What do you guys think about this?" Then I tweet it out, and then I get 10 people saying, "I want this," and then I can enable it. It's, it's l- from idea to deployed

34:00Thorsten Ballversion, it's really, it's short. Like, the biggest blocker right now, I would say, is our deployment pipeline. That's still, because we have, our scale is now large, so it takes a while to roll that stuff out safely to a bunch of these, um, processes running. But-

34:18Quinn SlackUsually like 10 minutes, though.

34:20Thorsten BallWell,

34:21Thorsten Ballthat-

34:22Quinn SlackThere are, yeah, there's some stuff that takes a few hours

34:23Thorsten Ball... for other stuff. For other stuff. Like I-

34:25Quinn SlackYeah

34:25Thorsten Ball... example yesterday, like again, like shout out to Ian. He was like, "Hey, can you build this for me?" I send off the prompt. It did it, and then it was, like, my 8:30. I had to put the kids to bed and whatnot, and I was like, "I really wanna get..." Like, "I really wanna tell him it's done," but then I was like, "Okay,

34:42Thorsten Ballit's 8:30. Gotta put the kids to bed. I'm going to bed then." I, like, I don't wanna roll this, you know. Like, that was the biggest thing, but it was done done, you know? Like, it was not done done, but it was done. So yeah, that's, that's a realization that Amp for us is what we previously called Jellyware, right? It's,

34:59Thorsten Ballit's high quality, it's high taste, and I'm quoting here, right? I'm not bragging. And, and yet we don't develop in a traditional sense of how other people develop software. Like, there's a lot of freedom and a lot of trust in this team on how and when people ship stuff. And it's not-

35:18Quinn SlackYeah

35:19Thorsten Ball... vibe coded. Like, it's not, it's not slop. And-

35:22Quinn SlackBut it's like when we use Amp, we are like homeowners. And-

35:27Thorsten BallYeah

35:27Quinn Slack... with homeowners, you can make your house your own. You can put whatever on the walls. You can paint it however you want. You gotta do all the maintenance, and if you see something that's wrong, no one else is gonna fix it. You gotta go fix it yourself.

35:39Thorsten BallYeah.

35:39Quinn SlackAnd that's how we feel, and there's a certain joy in that responsibility. And then the other way of using software is where you're just an end user, and you cannot go change it. When you're, you know, renting an apartment.

35:52Thorsten BallYeah. Yeah.

35:52Quinn SlackAnd it's nice that you can call the landlord, and they can go fix the washer and dryer, and you don't have to deal with that. But hey, you know, there's a lot of stuff you'd like to change about it, and you can't.

36:01Thorsten BallYeah.

36:01Quinn SlackAnd there's pros and cons to both. But all of a sudden, it's become so much easier to go and change the house and so much cheaper.

36:10Thorsten BallYeah

36:10Quinn Slack... and now, you know, a lot more people understandably wanna be homeowners. And so we're thinking about, for us, how we can get that AMP way to more people.

36:23Thorsten BallYeah.

36:23Quinn SlackAnd how we can give that to you in AMP, and also how we can help you get that in more of the software that you're using.

36:30Thorsten BallYeah.

36:31Quinn SlackAnd it does require change in how companies work. Like if we, instead of building AMP the way that we do, which is small team, a ton of ownership, the ability to fix things, to, to change things, not a lot of process. If instead we had a very slow issue product management roadmap process, and then we had a very slow code

36:53Quinn Slackreview and deployment and approval process, well then it, it wouldn't matter that we have the ability to change stuff at AMP because we, we just wouldn't be able to. We'd be bottlenecked on either side. So-

37:04Thorsten BallYeah

37:05Quinn Slack... we wanna, you know, as one thing, export the AMP way to more teams, share more about what that means and why the joy of being able to push to main and, and why that can lead to higher quality, and what you need to do differently on your team to make it so you can do that in a high-quality way.

37:22Thorsten BallYeah. That's what lies beyond the orbs, right? Like, what's the next steps? Like that's... Yeah, like orbs are just the next, it's the next stepping stone, right, to, to fully leaning into this. And it's funny because you-

37:37Quinn SlackAnd by the way, someone who has terrible process on either side, for them, if it takes you a week to ship something and three weeks to figure out what to ship, then an EC2 cloud with a bunch of shell scripts is gonna be just as good as AMP.

37:53Thorsten BallYeah.

37:53Quinn SlackNeither of them are they gonna be that good.

37:55Thorsten BallYeah.

37:55Quinn SlackAMP really shines when you can move so fast. That's when the multiplayer stuff, the portals, the orbs, the ability to do so much at once, the ability to have it monitor into production because it pushed and it is rolling it out. All of that stuff, we built AMP for that. So-

38:09Thorsten BallYeah

38:10Quinn Slack... you know, also-

38:11Thorsten BallIt-

38:11Quinn Slack... us exporting the AMP way will help more people see why orbs are different.

38:15Thorsten BallIt's n- only because you just brought it up, I just realized, like, how seldom I now think about these companies with traditional processes, you know? Where we get into contact with them when they say like, "Can you make orbs work for us with, in this workflow, in this environment?" And man, like, I guess like every software

38:36Thorsten Ballcompany should throw away,

38:39Thorsten Balleven as a thought experiment, how they've done things and then restart from scratch and think like, "Okay, if we take the trajectories that we just talked about, intelligence will-- prices of intelligence will go down, speed of tokens will go up, the intelligence itself will get smarter. If we take these trajectories and

38:57Thorsten Balltake them as a given, how would we develop software?" And I think a lot, the outcomes would look completely different. That's a whole, that's a summary of this conversation, right? But the funny thing is, the interesting thing to me is when we see companies, uh, you ment- you know, we mentioned Linear, um, two others come

39:16Thorsten Ballto mind, like Raycast, Shopify, for example. They're leaning heavily into... They were pre, you know, founded pre-2025, um, even maybe 2020 indic- I don't know with Raycast. But like pre-agentic companies, and they're leaning heavily into AI. And so you see them how reshaping the company to this. But the, the weird t- thing

39:40Thorsten Ballhere is that, you know, it's selection bias because we don't ever hear about the companies that don't reinvent themselves. Like they just die a quiet death. Like I'm not calling out a name, but I found a company, I posted it in our Slack.

39:56Thorsten BallMaybe I can say it's about email, but it was like the website looked like 20, you know, 2004, like 2008 maybe. And it's like, that's not... Like who are you select- you're selecting for people that use their computer like it's 2008, and you...

40:17Thorsten BallIt's not like they will go out with a big bang. They will go out with a whimper because the world has moved on, and it's just not gonna be interesting. So I don't know where I'm going with this. I'm just saying like-

40:28Quinn SlackYeah

40:28Thorsten Ball... you need to rethink how you sell them.

40:29Quinn SlackWell, it's easier said than done.

40:30Thorsten BallYeah, it's easier said than done. True, true.

40:31Quinn SlackJust, "Hey, everyone in the world, go rethink everything." But-

40:34Thorsten BallYeah, that's true

40:35Quinn Slack... that's why as we are thinking about, given that we have AMP, how can we have the biggest impact?

40:43Thorsten BallYeah.

40:43Quinn SlackAnd to be clear, it's not about us suddenly making AMP bend over backwards to accommodate these old workflows. No, we want to evangelize this new way of working and also take the fact that a lot of the software that these companies use is, you can now replace it with something that has 10% of the features, no settings screen,

41:05Quinn Slackis maybe 1% of the cost, and is infinitely more customizable.

41:09Thorsten BallYeah.

41:09Quinn SlackAnd you don't have to be a company that is ready to be agent pilled. You don't have to be someone who's using the agents directly in all the ways that we might want you to in order to get a ton of value from that.

41:21Thorsten BallYeah.

41:21Quinn SlackBecause you can go and say, "I used to have this CRM where, you know, one really smart person on my team had to manage it basically, you know, all the time. It was super clunky, it was super slow. And now I can just say, 'Hey, I want you to connect to Meta and pull in the data from here for our ad campaign.'" And that is

41:41Quinn Slacksomething that anyone can appreciate.

41:46Thorsten BallYeah.

41:46Quinn SlackSo it's how can we get this impact out more? And that's why I don't think that on the margin, the, you know, 100th person building their own agent platform product, I don't think that's gonna move the needle. I think it's time for us at AMP and really software developers to think about how you can actually go solve the

42:05Quinn Slackproblem, not build software to build software to, to solve the problem or something like that.

42:10Thorsten BallYeah. And you just reminded me of something or made it click, which is when y- you know, in that triangle of build versus buy versus build and buy, I don't know what the ... But there's a thing here where what you just said, right? Like, it's easier said than done rethinking your business constantly while the models are

42:29Thorsten Ballgetting better every two months, right? And while the costs are changing, and while the, the, the, the, how good they can do different things is constantly changing. So this is what we're doing at Amp. Like, if you're using Amp and paying us, you're literally paying us to constantly rethink how software is being done. Like,

42:48Thorsten Balland this is what I do all day long, right? Like, and that's in the product. Like, the way we think about how software can be built nowadays, that's the product. That's an expression. And it'll will, it will look different in two months, right?

43:00Quinn SlackYeah.

43:00Thorsten BallSo that's another thing. So in times of these technological changes, it's not just like a static build versus buy because the landscape is completely changing, right? It's like, d- you know, do I make my own rain jacket or do I buy a rain jacket? No, no, no. It's like you're going in the jungle and there's, like, a guide

43:20Thorsten Ballthere, and he's gonna give you the clothes for the day because he can read the weather there. It, it, it doesn't matter whether you buy or build your own rain jacket. The weather's gonna change, and you want the guy in there who knows, like, can read the clouds and whatnot, right? Not to oversell it, but I think that's-

43:37Quinn SlackYeah.

43:38Thorsten BallI don't know what that analogy is, but, uh, you know, like-

43:41Quinn SlackYeah

43:41Thorsten Ball... AI land is a jungle.

43:42Quinn SlackWell, I think if you import that to software, it's, it ... In the past, software was boring. It had to be mediocre. Because of that principle of the-

43:50Thorsten BallYeah

43:50Quinn Slack... software business, you build it in year one, and then you milk it for the next 10 years by-

43:54Thorsten BallYes

43:54Quinn Slack... because it's zero marginal cost copying. You know, you want it to appeal to the broadest market possible. And now that's changing. Like, why do I use-

44:04Thorsten BallYeah

44:04Quinn Slack... Ghosty? Because I trust Mitchell in-

44:07Thorsten BallYeah

44:07Quinn Slack... what he thinks a terminal emulator should be. I wanna see what is the issue tracker not made by a company that has to water it down to appeal to a lot of people, but what is the issue tracker made by someone that I really respect, that has crazy ideas, who's gonna show me not-

44:26Thorsten BallYeah

44:26Quinn Slack... just, oh, it works, but how I should be doing this in the future.

44:30Thorsten BallYeah.

44:30Quinn SlackOr how we do video calls, how we plan our team events, how we do travel, all these. And so in the past, if you had a great idea here, you would have to go and go hire a bunch of people, spend all your time doing that, not on the product. You'd go raise VC money.

44:47Thorsten BallYeah.

44:47Quinn SlackAnd in the end, the incentives would not be to build the most opinionated, correct product, and there would only be, like, a few competitors now. There can be-

44:55Thorsten BallYeah

44:55Quinn Slack... so many more competitors. And you could even have, I want to use so-and-so famous developer's or famous product person's product because their name is what gives me trust in it being different. And they can make so much more money, and it can be cheaper because there's so much less overhead that they need to go and

45:14Quinn Slackbuild that-

45:15Thorsten BallYeah

45:15Quinn Slack... and distribute that and sell that. They do not need to make a settings screen. They do not need to make it multi-tenant. They do not need to do all of this other stuff that they used to have to do, and that is really exciting. So in the same way-

45:28Thorsten BallYeah

45:28Quinn Slack... that we want Amp to be a representation of how we think you should be building, and if you trust us, then use Amp and we think you'll have a great time, I want that for all the software that I use.

45:38Thorsten BallYes.

45:39Quinn SlackYou know?

45:39Thorsten BallIt's, y- you're paying somebody else to think about the problem constantly. Like, that's what you want. I want Mitchell-

45:45Quinn SlackYeah

45:45Thorsten Ball... out there thinking about terminals constantly. I want whoever the person is at Google Meet who thought of when I switch tabs, they show me the little pop-up, you know, where the Google Meet meeting goes on, great job. Like, I want you to think about this. I don't wanna think about it. Which goes back to your point of,

45:59Thorsten Balllike, you want to do the thing and not build the thing to do the thing. And yes, intelligence is cheap and tokens are cheap, and does labor, like intelligence labor, is cheap. But, like, the thinking that a human does about a problem space is still different, right? Because it's continuous, and they have different options.

46:19Thorsten BallThey weigh different options if they're human. So that's kind of what you, what, what's worth it, right? That's what you're paying for. I want somebody, I want, like, a Rasmus Andersson designed text editor or whatever because I don't wanna think through the edge cases or Zed or whatever. Like, I can whip up my own, but

46:36Thorsten Ballfor certain things I want other people to think about it.

46:40Quinn SlackYeah.

46:40Thorsten BallBut yeah. I think we should, we should end it ... I'll, I'll end it with a short anecdote because you brought up Ghosty and the terminal. So we're all, we all killed our local dev, uh, uh, I don't know, barely use it anymore, right? And, um, I posted an outside, I said, like, "In the last six weeks I forgot all of the jujitsu

46:58Thorsten Ballor JJ shortcuts that I set up, the aliases. I was completely lost." And then Tim Culverhouse replied, and Tim Culverhouse, for those who don't know, he built his own terminal emulator. He built his own, I think, Wayland compositor window manager. He built his own email client. He built, like everything he uses, he built

47:19Thorsten Ballhimself. Uh, he owns Shell. He has its own Shell running in his own terminal. And he replied saying, "I have my, I have the best terminal and the best Shell in the world, and I don't use them." G- and it's, and a sad smiley because he's fully in Orbs now. So-

47:35Quinn SlackYeah

47:35Thorsten Ball... what does that say about, you know, the value of software?

47:39Quinn SlackYeah. And everything is changing faster, so if you are someone who has thought X and then you found X to not be true, I mean, we have, and it's very humbling, then you just get your mind to a place where you permanently are more open of crazy changes out there.

47:56Thorsten BallYeah. Wild times. All right. We'll see you next time.