Features Are Meaningless Now

Raising an Agent · Season 2 · Episode 8

November 18, 2026

Quinn and Thorsten talk through two launches, Spaces and Boxes, and what they say about building software when features can be dictated into existence. They explain why they shipped Spaces before they'd figured it out, why a list of features is no longer a reason to choose a product, and how the Amp sidebar changed all day. They also cover breaking changes you fix by sending customers a prompt, why self-hosted Orbs are free with no "contact us" form, Boxes as always-on computers with an agent running in production, rebuilding their bug tracker on a Box, letting your agent ask another company's agent, the permissions mess no agent could prevent, and whether to kill the settings screen.

Listen on other platforms

Transcript

Thorsten: When you can dictate features into existence, the value of features itself as a thing that your software just bundles devalues the software. Or it kind of doesn't make sense to me. Like, software should be a different thing now.

Quinn: I think that this applies to anyone who's building software. Features are completely meaningless as a reason why people would use your product, because an agent in general, what it does is it makes it so easy to add new features. And if you are ever asked, ‘Why is your product better than your competitor's product?’, a lot of times people are going to press you to give a list of features. Like, what are the things that you can build or that you have that your competitors will never be able to build? And in the past, the classic, the response would be something like, ‘Well, we have better integrations.’ Then the other person would say, ‘Well, why can't, you know, so-and-so just build that?’ And the answer is, well, now they can.

Quinn: And it's never going to be, ‘Oh, I built all of what your software does from scratch.’ It's going to be, ‘I built the 2% that I need, and I built the other 2% that you're never going to build because it's so custom to me.’

Thorsten: Welcome back to Raising an Agent. So, there's been a lot of stuff we've released in the last two weeks, and especially two things that we released, people really want to know how do we in the Amp team use them? So let's start with what we released last week, Spaces. How do you think about Spaces?

Quinn: Spaces, it's a way to stay focused on one thing you're doing. When it's so easy to kick off like 50 different things sometimes, and then you're bopping between, and you feel super productive. It's kind of like three years ago when you felt really productive in Slack where you'd go and you'd reply to everyone on Slack, and then you'd realize, ‘Oh wow, it's already 3:00 p.m. today.’ And that sucks. And it sucks because, you know, there's a list and stuff that wants your attention, and human brains, I mean, who knew when we evolved so long ago that we would be so susceptible to these kind of screens with lists, but we are.

Quinn: And we want Spaces to be a way where you can say, ‘I want to work on something bigger. I have an intent, a goal, and I don't want to be taken out of focus by all these other things going on. I want to just see these things, just work on this thing, bring other people in as well, and keep all of the work about a certain thing in a certain place.’ And it lets you also tackle things that are bigger than just a thread. It has really nice ways for it to run multiple threads, for you to see what's, you know, going, for it to have artifacts, and so on.

Quinn: But really it's all about giving you that focus back and making you feel calm and at ease when you're working with your agent. I don't think we have gotten totally there, but it's a huge step forward for me personally. And I was pretty skeptical originally when it came out. There's some people on the team that loved it, and for me it was really only in the last week or so, just, you know, a few days before we shipped it, that it really clicked for me. What about you?

Thorsten: Yeah. It's funny that you basically approach this from a whole different angle. Your angle is how it should feel like to work in this. And I think my angle is more, you have one main agent and it can spawn different subagents and to use Lewis's words, you sit in the captain's chair talking to the main agent with a bird's-eye view, and then you can dive into the subthreads. And I would say it's the result of two things, two observations. One is who cares about like the details of the thread? Like that's been the thing of the last six months, right? Where, think back, last year people wanted to see every tool call and whatnot. And now you just want to see your prompt and what the end result is, and then your prompt and then the end result. And everything else can be truncated. It feels silly to give so much importance to a single thread when it's basically just inputs and outputs that you want to look at constantly.

Thorsten: Two other things. One is I found myself often working on a cluster of ideas. Say I'm working on, I don't know, a new CLI, I can see how I will work on multiple things in parallel. I will start the news post, I will start the CLI, I will start the docs page, or something like this. I can start these in parallel. Before Spaces, I was the person, the meat proxy, who then had to coordinate between these three threads and like, ‘Oh, news post, by the way, check in with that other thread what the flags are,’ and, you know, ‘Also send an update to the docs thread and make sure that this is updated.’ And I found myself doing this. And actually, when I started working on Boxes, which we'll get to in a second, that's what I did, because we didn't have Spaces yet, and it felt pretty dumb that I was copy and pasting stuff in between.

Thorsten: So Spaces to me is like—you know, like in a lot of software, it's—what do you call this? Like a manifestation of a pattern or something, like a thing that's now in the product, a design that's now in the product that kind of encodes this pattern of you have a single topic, a single strand, where you have multiple things working at the same time and you want to coordinate. And that's what Spaces do for you, I develop Boxes, again, talk about in a second, in a single thread then. I think it's 180 threads in one Space. And I would fan them out, and then coordinate, and it would do the things like, ‘Oh, I noticed this one thread, it changed authentication. It also needs to know about this and update the docs.’ And it would do this for me. And it was nice just having steady set of contexts in one main agent thing. So, that's how I approached it. It's less about like, ‘Oh, take all of the notifications away.’ It's more centralizing contexts or sharing contexts between things.

Thorsten: But to address the elephant in the room, if people who are listening also have used Spaces or have used Amp, the fact that we answer with two different answers kind of shows that, we still haven't fully figured out Spaces yet. If you were to ask the whole team, ‘How do you use Spaces?’, I don't think you will get one single answer. I think you will get five, or six, or seven. I don't think that's a bad thing. It sounds like a bad thing where, ‘Oh, you don't know how to use your own product yourself?’ But I do think…

Quinn: Or you don't get aligned on what the goal is. You don't know who you're building for. Is it just going to get piled up with new features and actually not be good for anyone, right?

Thorsten: Exactly. But I do think—so we made the call to release this because we've been internally experimenting with this and playing around with it, and Lewis has basically built the whole thing, Lewis, and Tim, and Niko. It felt exciting watching this. It felt exciting trying it out. I said like, ‘Let's get this out, even if we're not 100% sure,’ because we will gain so much from hearing how our customers and users use this. And it's also additional. You don't have to use it if you don't want to. So it's not like we're breaking anything. So the risk is low. But if people use it, we get feedback, and then we can iterate. And, it's no longer 2021 where you can say, ‘We're going to polish this and work on this for the next three months. And then here's the release date.’ And we set the release date in August, and then at the end of November it's going to come out and we have everything figured out and all of the docs and the guides, and the DevRel people and blah. Everybody knows this. It just seems insane to me to take this approach in 2026 with AI. It's…

Quinn: Yeah. Also, the funny thing is, I think we all know that so often when there is some big launch and everyone celebrated getting that out the door, then, you know, some might say, to get the promotion, then all the good people go on to the next new project and no one follows through. So, it's kind of funny. Maybe it was very polished on the day that it was launched, but then it failed to keep up. And now if you fail to keep up, if you fail to follow through, then, you know—well, you shouldn't, because it's a lot easier to follow through now, but, you know, the thing's going to decay in like a week or two, rather than in, I don't know, maybe three or six months, like in the past.

Thorsten: And it also feels like that software was made of a different material even. Like, Spaces, it's ultimately a thing around the model. So if tomorrow—I we shipped Spaces last week, then Fable 5.2 came out this week, right? What if next week the next GPT model comes out? Maybe it will be completely different as a coordinator or something. How do you polish something for three months You have this cadence of new model drops that come out and completely change the game again. And at the same time, your customers are jumping from this model to this model, re-changing how they work. People send us messages, ‘I stopped doing PRs. We do trunk-based development because of you guys.’ And it's like everything right now is moving and shaking, and it just feels like you got to get stuff out, you got to hit reality, you got to get the feedback in, and then iterate on it. You know, obviously there's risks, and I tweeted this out yesterday to a user who was saying, ‘Oh, it feels like, you know, blah blah,’ somewhere on Spaces, and I said exactly this. I said like, ‘Look, two options: either we keep it under wraps and keep polishing it,’ which seems dumb, and insane for all of the reasons I just mentioned.

Thorsten: ‘Or we get it out in the open, and we answer questions, we can say or have to say, "I don't know. We're not sure. What do you think?" But then live from the feedback.’ And that person replied back and was like, ‘Yeah, I think you made the right call. Like, yeah, I can see you're changing it, but I think that's good.’ So they appreciate it.

Quinn: Yeah. I think we have the best users in the world. It's people who want to be walking on the frontier with us, who want things to change, who are okay if there's an abrupt change in their workflow and hopefully we get it right, but hopefully we've earned the trust to do that. And I think generally, I mean, the product's obviously gotten a lot better in the last month or two months or three months.

Thorsten: The…

Thorsten: Yeah. So there's this funny thing. I want to hear your thoughts, whether you notice this, too. So, obviously, we have users and customers who've been with us since, you know, one and a half years now. And when we do something, it's like they're part of the team, like they immediately get our stance and they're like, ‘Hey guys, what do you think about this?’ And, you know, they're fully bought into the whole thing. But now we do have the ability for customers to create their own shared Slack channel with us, right? If you're in the workspace and you have a certain amount of users and you have a certain amount of spend, which is not that much, you can create a shared Slack channel. And you can notice when a company basically has not been with us or used Amp for a while, where in their Slack it's like, ‘Guys, did you change this again? Or how should I use this? Like, what do the docs say? Don't you use this?’ And it's not like, oh, they don't know the product, it's very much that our customers who've been with us longer, they buy into the whole, ‘Of course you're going to change this. Of course you're going to rip this out. Yes, guys, if you rip this out, then I trust you. It probably doesn't make sense.’

Thorsten: And then new users have to get used to this, like that we ship five updates a day, you know?

Quinn: Yeah, but we do listen to feedback. Like, I changed something. I made it so that even if a thread was waiting on like a GitHub webhook, it would still pulse. And a lot of people complained. I had a really good idea with that, but it ended up being totally wrong and I changed it. So, keep the feedback coming. And that's exactly why we get these things out. It's our best idea and it's when we feel proud about it, that doesn't mean it's perfect, that doesn't mean it's done. And you get to co-create all of this with us because we're all figuring this out together. I mean, if your workflow changes in a two-week period because now you're using something like Spaces, then you're going to have all kinds of new demands for how Spaces should work. And when we shipped Spaces, there were maybe three or four people in the world that had used it for more than a week. And there's no way that we could have learned everything there. So, keep the feedback coming. You all have a huge influence on what we do. And one thing that I love is when people send us a video of them just clicking through the product and showing what they do and talking about their pain, I love it because it is so much information.

Quinn: You know, in the same way that dictating to your agent is a really fast way to get bits from your brain to the agent, watching someone actually click through and talk through and narrate the problems that they have, it helps us understand it so much better. It helps the agent go and make Amp better for them. So, send more of those. And maybe the bug report should default to that. That would be kind of funny. It'd be cool to see the, you know, faces and hear from everyone.

Thorsten: Yeah, Puck is going to show up and it's going to ask you, ‘Explain it to me. Can you show me, you know, what goes wrong?’

Quinn: Like one of those dystopian AI interviewers. Oh, man.

Thorsten: Yeah. But, okay, well, a little tangent here. You know, customers, some of them or a lot of them, our customers write amazing bug reports, and it's just a pleasure. I mean, sometimes they use Amp to debug Amp itself, so they get like diagnostic data, they have logs, they send us the thing. I put it in my Amp. We're both meat proxies at this point. I fix it. I tell them, ‘Yeah, it's fixed,’ and whatnot. And that's great. But man, side note, you know how people lament like the fall of the craft of software and like the loss of the craft of software and whatnot? Dude, if you think that the craft of software was something widespread and just not a niche thing, I urge you, go and look at the bug tracker of an open source project. It's insane. Like, people do not know how to use software, and—it's developers. Like, it's developers leaving these issues on, you know, bug trackers of like developer products, and it's insane to me. So…

Quinn: There's a few categories. It's either like a total slop thing and then usually it's like, you know, step one, fix the bug.

Thorsten: That…

Quinn: Step two, ensure that there's no more bugs like this that exist. Step three, make a better process for—and it's like so slop. And then sometimes people are like trying to play like 20 questions and they're like, ‘Oh, well, you should know. Isn't it obvious?’ And, you know, ‘Oh, well, this other tool has it.’ And the funny thing is, a lot of times, they literally started using that other tool, if you go on their, you know, posting history, like five days ago. And all of a sudden, it is the shit to them. And then every other tool they are looking at must have all those features. But then why are they switching if they're so happy? And that's where it's really helpful to try to earn users' trust in a bigger way, earn their team's trust, because then we know that people who are sharing this feedback, they have skin in the game. They're not just saying, you know, ‘Oh, I want all the other features—I want the superset, you know, of all the possible features, and then I'll try Amp.’ And obviously, and this gets back to the motivation for Spaces, if we did that, our product would be a clown car, and it would be incredibly distracting, incredibly complex, and then you wouldn't like it.

Quinn: So, what we love is hearing from people that talk about what they actually need and not like, ‘Oh, this other thing has that, so I want it.’

Thorsten: That—I don't get this. Like, I don't get it, like why—it's like you switch from Emacs to Vim and then it's like, ‘Well, you know, my left thumb feels too good. Why don't you add like chording and,’ you know. I don't know why I'm complaining, but it feels weird when people switch to Amp and then they say, ‘Add all of these other features from these other products.’ I don't know. Like, is that valuable that you have all of the features from every other product just in one box? Like, I don't know. Like, I don't think it's valuable. I don't think in the times of AI that having just the most number of features is valuable. I think, I've said this before, that people, when they sign up for Amp, it's not that they sign up for like a list of features or a matrix of features. What they sign up for is to unleash these models to their full potential and work in a way that lets you get your stuff done and get value out of these models. And also, you sign up for that being on the frontier of what's possible continuously, you know? You're not signing up for a snapshot of a list of features. You're signing up for a way of working and getting the most out of these models.

Thorsten: And if we had done the whole build every feature, then by now we would have had like a PRD-based workflow. We would have had like Ralph loop first-class support. We would have had /btw. We would have had fucking to-do lists still. We would have had pets? You know, like

Quinn: Tab completion and an IDE, we'd have that.

Thorsten: Tab completion would still be there.

Quinn: We'd have a really good interface.

Thorsten: It…

Quinn: We could use JEV to figure out should we apply this part of the hunk of this diff, you know?

Thorsten: Yeah.

Thorsten: It doesn't make sense. I don't know. The larger point to me is when you can dictate features into existence, the value of features itself as a thing that your software just bundles devalues the software. Or it kind of doesn't make sense to me. Like, software should be a different thing now. We don't know yet what it is, but it should be a different thing.

Quinn: Yeah. I think that this applies to anyone who's building software. Features are completely meaningless as a reason why people would use your product, because a coding agent, an agent in general, what it does is it makes it so easy to add new features. And so there's really no barrier to you adding a feature. I mean, I'm sure there's some companies where there's a barrier, but they have even bigger problems.

Quinn: And if you are ever asked, ‘Why is your product better than your competitor's product?’, a lot of times people are going to press you to give a list of features. Like, what are the things that you can build or that you have that your competitors will never be able to build? And in the past, the classic, the response would be something like, ‘Well, we have better integrations and, you know, we have blah blah.’ And then the other person would say, ‘Well, why can't, you know, so-and-so just build that?’ And the answer is, well, now they can.

Quinn: And so what we're facing in this incredibly, competitive, dynamic, fast-moving agent market, I think it's what everyone else is going to be facing in every other part of the market. And then we're also facing people who say, ‘Well, we have built an agent in our own company,’ or, ‘I have my own agent,’ or, ‘I've made one with Pi,’ or something like that. And I think that other kinds of software, they're going to be facing that, too. And it's never going to be, ‘Oh, I built all of what your software does from scratch.’ It's going to be, ‘I built the 2% that I need, and I built the other 2% that you're never going to build because it's so custom to me.’

Quinn: And that's where it's really nice that we can just adopt a philosophy of like, you know, ‘Great. We are at peace with that.’ It's a wide world out there. The scope of an agent is so large. This is not like an HR system. This is not like a screenshot app. The scope of an agent is everything that you can do with a computer and with your team and in communicating with the outside world. And so people are going to have such a wide array of preferences about it. And it's so undiscovered, so people are also trying all kinds of things to figure it out.

Thorsten: Yeah.

Quinn: So that's just to say that it's not about what features we have or don't have. You know, anyone can add new features. I think it's about if you are going to use a product and learn, and does your workflow going to get better and better? And I don't think you want to be switching from product to product every single week. Some people I think enjoy that, it seems. But if you don't enjoy that, then, you know, find a team that you trust and really, you know, dig deep with them.

Thorsten: Yeah.

Thorsten: Let's talk about something that, you know, it's adjacent to this, the sidebar. So, the sidebar changed with the introduction of Spaces. So, now you had Spaces and Threads. That was the first thing. That was last week.

Thorsten: On Monday, in the meeting, we said, ‘Amp should feel calm when you work in it.’ Like that should be the goal. It shouldn't be, you know, popping up with notifications.

Thorsten: Then, yesterday, I lost it and I wrote in Slack like, ‘We need to fix this ASAP. I got 50 notifications everywhere, there's things popping up and it feels overwhelming.’

Thorsten: And then you basically pushed the eject button and built a new sidebar, or basically re-changed the sidebar that we had in the experimental settings and changed it completely, and it changed over the course of the day yesterday, while I was using it, from like my 6:00 a.m. to my, I don't know, 8:00 p.m. or something, while I was always using it. You pushed a bunch of updates, others pushed a bunch of updates, and it kept evolving throughout the day. Obviously like a, trite way to say it would be like, ‘Oh yeah, that's evolving software, and it's, you know, whatever. There's a bunch of updates.’ But I do think there's something else going on, where you throw a new thing into this and then you make it experimental, and then we have our users immediately—there's people constantly going through our experimental settings and looking for new stuff to try out, and they immediately hopped on it. So, yesterday, while you were working on it, while I was using it, we had people on X and in Slack messaging us, ‘Oh, the latest update is good. The latest version is good.’ And it felt so weird and strange and alive that it changes.

Thorsten: And then yeah, it's cool.

Quinn: It's so cool.

Thorsten: It's really cool, but it's wild, and like, I don't know, like, yeah, it's wild and cool. I don't know how else to put it.

Quinn: I think there's a lot of people out there that love having people use what they build. And it's also really nice if the time between when you build it and when you get feedback—good or bad feedback—is short. That's a really nice feedback cycle.

Quinn: And that's what we had with the sidebar. That's what anyone can have. But you got to get your company into a place where you can configure it. So I think it's interesting to call out the things that just came naturally to us where we just thought, ‘Of course we're going to build it like that in the open.’ I think that it's a certain level of trust that we've earned with our users. I think it's a risk tolerance on our side that there were some cases where I didn't think of all the concepts there. I think it's a very light touch ownership model in Amp where who owns the sidebar? Nobody. And now you could say, ‘Well, that's why it sucks so much. That's why it is so distracting.’

Quinn: And that might be right, you know, a little bit.

Thorsten: Yeah.

Thorsten: Yeah, I want to explain how it happened, right? So, in Monday's meeting, because people are always interested in this, in Monday's meeting, we were talking about this issue, of—it wasn't that big of a problem, but it was this like, ‘Oh, we need to figure out.’ I said, ‘The biggest thing we have to figure out right now: agent-to-human interaction points, agent-to-agent interaction points, and how they tie into Slack, in Puck, in Spaces, and in Conversations. We've got to figure this stuff out. You know, like, you shouldn't feel confused about pasting stuff in windows and not knowing what's being sent where and why.’

Thorsten: That's kind of what I said, and then we talked about it and everybody agreed, and then Tim Lucas was saying, ‘Yeah, Amp should feel calm to work in.’ And then, when we got to the planning for the week, who's doing what, he basically just wrote down, ‘Spaces,’ he's going to work on Spaces, and ‘Amp should feel calm.’

Quinn: And he…

Thorsten: Then, I think a day later, he added a new experimental sidebar and he wrote in our work-in-progress channel. He's like, ‘Hey guys, enabled a new sidebar.’ I didn't even know that he's going to work on the sidebar. I assumed he was going to work on different things. It was just his approach was, ‘I'm going to work on this.’ He didn't ask us for—you know, doesn't have to ask us for permissions, but he didn't like share plans of what he wants to do. He's like, ‘My goal is to make Amp feel calm.’ So, in his mind, the solution is, ‘I'm going to change the sidebar first. I'm going to enable it for the team and get feedback.’ He enabled it.

Quinn: And you know what? During that, I was making some changes to the sidebar as well, 'cause I wanted to fix it. We did not coordinate, and, yeah, he didn't check with me. ‘What? You know, he's stepping all over my toes.’ No, he had great ideas and he was in a place where he was like following his energy and motivated and so was I. And it turns out that I think, you know, we both had some good ideas that came together, even though we didn't coordinate until after.

Thorsten: Yeah. Yeah.

Thorsten: Yeah, but then, after I saw you guys in Slack, he's like, ‘Hey, I shipped this,’ and then you were going in and saying, ‘Oh yeah, I also made these changes. Do you want to pull them in?’ And then he's like, ‘I'm going to bed,’ because he's in Australia, and then you're like, ‘Oh, I'll merge this in.’ And then, I was like, ‘Oh, I hate this. I don't like the sidebar.’ And then Tim Culverhouse was like, ‘I don't hate this.’ And then it went back and forth, and then yesterday, we reached this like, ‘Oh, we got to fix something,’ and then you went in hard, and then Tim went in hard, and now, I think it's in a good place. So, if you look at the last four days, it's better.

Quinn: It's better. There's still a lot of work to do.

Thorsten: Yeah, but by the time this recording comes out, maybe it's fully polished, but who knows?

Quinn: But…

Quinn: Yeah.

Quinn: I think it's so important to call out that there is no ownership of the sidebar and maybe that is why it got into a bad place in the first place, but the lack of ownership of the sidebar means that we can also just point a bazooka to go and fix it. And we don't need to wait for someone who—you know, if the ideal case is someone owns something and it's really good. But sometimes someone owns something and they have other things going on or they don't have the right idea or they're in a rut or something. And so whoever has energy from working on that thing can go after it.

Quinn: And when all the work is done in Amp, all the human intent that Tim ever had about why is this thing in this way, why are these nested here, why did he choose this instead of that, that's all recorded in Amp. So he was asleep and I was able to figure out what did he mean by this, why did he do this. And there's a lot of cases where the agent said, ‘Well, there's no actual human intent there. The agent made an assumption. It was never checked by the human.’ And then I was able to overturn that. And then I could also overturn things that he did by his intent. But I was obviously a lot more reluctant to do that because I thought, ‘Oh, well, he's thought that through.’

Quinn: And it's this really magical feeling, and you can make stuff better so much faster. And there's no way to do it if you have the traditional team structure or process or you have to wait for things to get merged, or if you haven't set expectations with the people using your product that they're part of a journey of exploration with you.

Thorsten: I think that's a really good point. I talked with Mats yesterday about this, that ownership obviously is a good thing. You know, like, this guy owns this and this person owns this, and we talked about this before in the last two episodes. But there's obviously also a downside to ownership. Like, if this part of the codebase has the label Quinn on it, then people might feel not encouraged to change it however they want and go in and make changes. In a small company where you have so much to do, it doesn't make sense to divide it into ownership, and it's just like, yeah, if you want to work on this, go and work on this. If you want to work on this, go and work on this. And the fact that you can just go in and change things, or raise questions even, that makes a lot of stuff more possible that you cannot do if you're in a bigger company where, you know—imagine I blew the gasket as I did yesterday, was like, ‘This fucking sucks. I hate this, and I hate looking at this.’ And then what? Imagine if we had a sidebar team. Imagine this. Imagine that I then had to go and talk to the sidebar team who's been doing something else, and they might not work on the sidebar, but the next version of the sidebar, and then you have to coordinate with them.

Thorsten: Obviously there's benefits to teams and having more people, blah, but what I'm trying to say is you should use the fact that you don't have these things. If you're a small company, there's advantages to not having these, and you should try and benefit from them. And it's not obvious because to everybody in startups or in software, the best practices are always based on, big companies and how the big teams do it. And like, everything else is a little bit toy-like, but there's a lot of stuff that goes well when it's just you and your buddy alone in a project. There's a lot of stuff that's good, and you should try and benefit from this. And it's not like a toy-like thing, or it's not something to smile at. It's a really true thing that you have a huge advantage if it's just a small number of people working in a large space. They have a lot of autonomy, a lot of room to breathe, and they can do whatever the hell they want. So I—

Quinn: Yeah. I think the test for anyone listening is, can you point a bazooka at any part of your product, say, work on it for four hours, you can make something and with an agent so much better in four hours? And can you do that? Can you do that without asking anyone, telling anyone, getting permission, without ruffling any feathers, without upsetting the people that were depending on the other thing that you were working on, without needing review? Can you get it out to customers, even after two hours, and get some midpoint feedback? Can they use it?

Quinn: And if you can, that's awesome. And you feel the joy that we feel. If you can't, just know that, man, it feels so good to be able to do that. And start small. Pick like the dumbest, most boring part of your app, like, you know, the billing settings, and just make it really good. See if you can do that. No one's going to get that mad if you do that. Try it and report back.

Thorsten: Yeah.

Thorsten: I think there's another thing, just popped in my head. Maybe we'll go on a tangent, maybe not, but I think I have to bring this up.

Thorsten: People still don't take into account how much has changed due to the fact that you can now make major changes in a couple of hours with these agents. You know, it's mind-blowing.

Thorsten: So yesterday, I was talking at an all-hands of a different company, and one of the questions they prepared, and I didn't get to answer it, so I'm going to answer it here, was, and I want to hear your thoughts, the question was, ‘How do you handle, you know, moving so fast and shipping a bunch of stuff,’ and the sidebar is a good example, ‘how do you handle, when a customer has built something on your API? Like, if you have a fixed API, and customers depend on it, say schema or whatever, an API shape, how do you keep moving fast?’

Thorsten: And my answer is, obviously I don't want to break their stuff. But consider, what was always the pain in the ass of maintaining, backwards compatibility or, you know, breaking changes? It was maintaining backwards compatibility. That's a pain in the ass. Now with agents, it's pretty fucking trivial to say, ‘Keep these two things alive.’ So it's not that bad to have like another version of the API, a stable version, and then deprecate it a couple weeks later or something.

Thorsten: And at the same time, and that's the most counterintuitive thing, is if you now say to your customers, ‘Guys, we're going to break this,’ you're not making them jump through hoops. You can now give them a prompt, and say, ‘Give this to your agent. These are the changes we make in the API. It should be able to do this and migrate it over.’ So it's like you can send somebody from your team down in their team to make the migration for them, and all it costs them is a bunch of tokens. And for stuff like this, it's usually not crazy amounts. It's like 10 bucks or something.

Thorsten: So, that's like another thing where, this was a problem back when writing code was really hard. And now that I can send you a prompt and you execute it, a lot of stuff gets easier. Even breaking changes or moving fast gets, you know, easier with this.

Quinn: Do you think that we can say that because we're a coding agent that is explicitly for the people that want to be on the frontier with us and work like that, and it doesn't apply to any other coding agent or any other kind of software?

Thorsten: Yeah. I mean, to a certain extent, yes, but look. It's a good call, like we made a good choice. It's—

Thorsten: I mean, the whole game is changing, and obviously it's not like a one-day thing. So if the AI diffusion was at the max 100% tomorrow, then I think we would be even more right? But obviously the counterargument is, ‘Well, the customers I work with, they use like $10 a week, you know, whatever tokens, and they're not AI-pilled. I cannot give them a prompt. Their security would kill me, blah, blah.’

Thorsten: So there's obviously this argument of, ‘What you're talking about is a future that you're living in, but for 99% of the rest of the industry, it hasn't arrived yet, so it's not valuable.’ And yeah, might be, but you know, I don't know, what am I supposed to say? Like, yeah, our job is to be on the frontier of this, and I think you've got to get there, you know. It's the same with other technological changes, except that now the timeline is compressed. It's the same with, you know, people using the internet back in the day. Like if you had your own customers that you could talk to via email, you would get a lot more stuff done, and it would be easier. And then, yeah, one customer is like, ‘Oh, we still use, you know, mail, and you've got to send us a letter,’ and whatnot, then it's like, ‘Ugh, come on, man. You know? You know this is the future.’

Thorsten: So, I don't know. It's just compressed, but that's always been the case with technological changes. What do you think?

Quinn: Yeah. I think you're totally right that agents make a lot of other things about maintaining software, self-hosted software, things like that, so much easier. And self-hosted is something that we did at Sourcegraph. I think we did it better and a lot more volume, for a lot more customers than other companies out there, but we started to feel the pains of it. And we would have customers on versions that were two years old, and it was such a pain to maintain. It became a real headwind.

Quinn: But now, with an agent, you can ask it, ‘Who are all the people that are on older versions? What do we need to do to maintain a shim so that we can keep supporting them, and to show them a very contextual notice about what might change?’ We can email them and give them 14 days' notice, and give them $10 in free credits and say, ‘Thanks for being a great customer. Let me know if you need help with this.’ They can run it. If they have errors, then they can email back, and then we can look into that with our agent. We can do easy brownouts, which is, if you're deprecating and removing an API, you start by 10% of the time you return an error, and then that's like a little smoke test to see what starts to fail. And the error message can contain something, even in some cases, custom to that user that says exactly what they need to do to migrate. And they're going to feed it into their agent. And I've done that a lot, and I found that people just feed that into their agent, and then they stop calling that old API.

Quinn: And there's a little bit of that where it's like, yes, it does take up time, but it's an incredible feeling. And I think it feels a lot better for a customer than even like a, ‘Here's a six-month deprecation notice’ and all that, 'cause it's a lot less time that they spend. And ultimately, customers know that everything is changing around us, and they want to learn what is the better way to do it. They want us to be in a place where we can move fast.

Quinn: So, I think the really interesting thing is we're starting to reconsider some of the assumptions we had in Amp where we said it's going to be multi-tenant only. And right now the central Amp is still multi-tenant only. There's a single instance of it. But we've now made self-hosted Orbs. That worked really well. I think, you know, before we'd even shipped it, there were more than 100 people just like in, you know, a few hours found it out and set it up. Can you imagine like a hundred people in a timespan of a couple hours setting up like a Helm chart and Kubernetes and all that stuff for some app back in the day? No way.

Thorsten: I mean, I'm very proud of this choice with the self-hosted Orbs, that basically the way it works is, you know, in the olden days if you want to self-host Orbs? ‘Here's a YAML and you define the provider and then, based on the provider, we load the appropriate plug-in, so that we can use Vercel Sandboxes, Baseten, E2B, exe.dev, and whatever.’ And we would put out officially supported provider plug-ins that we vetted and that work and whatnot, so that people don't have to go and read the API docs of their preferred provider and like write up the script and the integration. Like, because writing code was hard. That's why we did this type of stuff.

Thorsten: With self-hosted Orbs, it basically says, ‘If you want to self-host Orbs, here's the contract, here's the script that you have to write and it has to answer to these three requests, spawn, delete, and resume, where you have to put in the code that matches your infrastructure.’ So, if you have a self-hosted say Docker setup, we say, ‘Yep. You want to self-host? Here's our piece of the API. That's your piece of the API. Go and ask the agent to figure out the rest.’

Thorsten: And we do provide some scripts in our docs, example scripts. And we did have people like ask us, ‘Oh, can you add this provider to the list of supported providers?’ And I told them, ‘These are not supported providers. It's just examples, because I expect you to write the thing yourself.’

Thorsten: And it was a bet. And then within the first few hours of self-hosted Orbs being released, people were saying, ‘Yep, I got it working with my own infrastructure.’ I'm like, ‘So, what did you do?’ ‘I just gave the agent the docs and it figured it out and had it running.’

Thorsten: And then Andrew from Amp Labs, he's like, ‘Oh, this is funny. I got it working with—my buddy built his own hypervisor VM management system and it works with this.’ And I'm like, ‘I have a guess, but tell me how you did it, how you made it work.’ And he's like, ‘I just gave the agent the docs and it built the whole thing. And it can, you know, feedback loop and whatnot.’

Thorsten: So, I think that was a great choice and it's leaning into the fact that code is now cheap to produce. Like, we're talking about 200 lines of, glue code basically. That's solved now. So, when we talk about self-hosted stuff and breaking changes and bringing users along, I think previously the bottlenecks were: it's hard to debug, it's hard to look at logs, it's hard to write shim code, it's hard to understand APIs, and none of these things are true anymore. So, a lot of stuff is more feasible now.

Quinn: Yeah.

Quinn: So, we're actually considering what else can we extract from the central, monolithic deployment of Amp and make it self-hostable? For example, could we make it so that you own the cold storage of your threads, if you want? And then we would only have the thread data in volatile memory and storage, for example. That's the kind of thing that would be the project that you would do when you were like a 5,000-person public company and you wanted to go start working for the US military or something, and you'd have a team of 200 people on it, and you would gate it behind like seven ‘contact us’ forms so that by the time someone even just saw that, they were paying you like millions of dollars. But this new model is different, and we made it, intentionally, so that self-hosted Orbs was totally free outside the enterprise, because, well, it doesn't cost us anything for those people to use it. We learn so much from the feedback that they bring us, and also, I think we all know it sucks if you have to go do like some ‘contact us’ form to try something out. And that friction, the humans hate it. The agents—I can't. Well, I guess they can contact us, and they do in great volume.

Thorsten: Yeah.

Quinn: And it's a totally different mentality.

Thorsten: So, yeah.

Quinn: But if you had a lot of salespeople where they saw the way that they were going to hit their number the next year as this super complex, you know, minimum $5 million a year feature that's for the only, you know, biggest enterprises, there's no way in hell they're going to let it show up on some dumb developer-looking docs page where just anyone can do it and whatever. And I'm sorry, but if other companies, if some of our competitors, choose to go that route, many of them do, then that makes our product better and their product worse, both because you got to go talk to their people to try it and because we have so much more feedback, it is tested in so many different configurations. And I just felt so at ease when we actually shipped self-hosted Orbs, knowing that like, you know, 100 or more people had already set it up, and I think we'd gotten like one or two pieces of feedback about it not working. It was so good, and I know—

Thorsten: It's truly mind-blowing. Like before we shipped it, basically you and I at this time of day, which is your night and my morning, we made the call of, ‘Let's just give it to everybody for free.’ And then right before we shared this in Slack, and then rightfully so, people shared their doubts or fears or anxieties about this move where, ‘Oh, now we got to support everybody and it's going to be a lot of bug reports and whatnot.’

Thorsten: And turns out, you know, there's not a lot of that. Like, the bug report—we need to talk about this. The bug reports come from something else. They come from other things, which is not something that you can solve by just throwing code at it. But, you know, integrating our pretty slim API with your own infrastructure, the agents are pretty good at it. So—okay. We've got to talk about this, like, what the agents not solve. The whole complexity thing, you know, whatever. But first, let's talk about this other thing that we also shipped. So, this was this week. We shipped Boxes this week.

Quinn: If you were to say what are Boxes in one word?

Thorsten: Computers.

Quinn: Two words?

Thorsten: Always-on computers.

Thorsten: How would you say? Always-on, that's hyphened? It counts as one word.

Quinn: I mean, in German, you could probably, you know, this would be very easy.

Thorsten: Yeah. Oh, no. No, you can't do it in—well, I guess you could. How would you describe it?

Quinn: Instant software?

Quinn: Like omnipresent software?

Thorsten: Interesting. Interesting.

Quinn: I love that I can tell Amp, ‘Go and build this thing,’ and now, Amp, instead of just being for building it can also run it. And now, when we say run it, you know, deploy. Deploy is such a heavy, official word.

Quinn: No, like deploy makes you think of a Kubernetes cluster, and scale, and web-scale, and microservices, and all that, but the vast majority of software that you need—

Quinn: And this is why you built Boxes in the way you did—

Quinn: Yeah, you know, even if you have 1,000 people at your company, what is the queries per second of that? How big of a box do you actually need? And so, Amp lets you actually run it, and lets you actually use it, and you can click on it, it's right there, and I think that it takes out this idea of deploy and all that completely.

Thorsten: I mean, that's so interesting that you start with instant software. I mean, that's how it started, right? Like, the background of Boxes is that we, and our customers, found ourselves constantly building small things in Orbs and one-off dashboards. Or not one-off, those stay in Orbs, but like multi-use dashboards or little apps. And then we wanted to have a way to deploy them or use them. And also, I mean, I'm speaking for myself, they should still be close to the agent. Like, most of the stuff that I build with agents, like small stuff for me, I don't want it to be somewhere else and running. Like, I don't need a Vercel or AWS Lambda or whatever, like a scaled, serverless type of thing. I just want like a URL where I can run the thing and my agent should have easy access to it.

Thorsten: And that's how Boxes came to be. And that's also the biggest influence on their design, which is they're not replacing your Kubernetes thing. They're not a new, I don't know, SaaS hosting platform. It's truly computers on which you can run software and there's an agent running on them. And the agent, yes, the agent runs in production. There's an agent in production that you can talk to. And it's marvelous. It's amazing. And I'm going on the record, the agent will probably make less mistakes than a human in production.

Thorsten: And the whole reason is that I have these tiny little things and I deploy them. And then I'm using them on my phone and I was like, ‘Ugh. I typed in the wrong thing,’ or ‘Now I have a screenshot of like 15 links that I want to add and now I got to copy them by hand. If I could just have the agent add them to the database for me, that would be amazing.’ And now I can do this.

Quinn: So, instead of having the agent build a nice settings UI and an admin panel in your application, you just have an agent running on that machine that goes and manually like edits the SQLite file that it's running from.

Thorsten: Yes. I mean, it's malleable software. That's the idea. I hate admin screens, I hate settings screens. It doesn't make sense to me to have a batch select of like these five posts, you know, move them into the draft state or whatever. When I can say to an agent, ‘The five posts, move them into draft.’ You know? Like a lot of stuff doesn't make sense anymore, and when you think about it, the other influences, a lot of software now doesn't have to scale to 1,000 users at the same time. A lot of software is made up on the spot, and yes, you know, if Ian Landsman is listening, yes, nobody's going to build their own Figma, I get it. But there's a lot of other stuff that people build, internal stuff, our customers do it, and that doesn't need to scale at all. And a SQLite file is amazing, you know? And like having a cloud database and your thing running in a serverless and then you have to figure out the credentials and, cannot actually talk to the application because the application is like five containers actually and you can't get to it and then the database is another thing and—that's the whole thing, and if it's just a computer, there's a SQLite file on there, there's a web server on there, it's amazing how much they can help you run this software.

Thorsten: They can look at logs, they can spawn Orbs to modify the software, they can change the database file and it's, you know, what we've been talking about, Jellyware. Jellyware is the original idea behind Boxes, and now with Boxes you can build a bunch of this stuff.

Quinn: It feels so wrong. Like everything we learned about having cattle, not pets, everything we learned about robust deployments, and everything should be in a cloud database that's backed up, nice admin panels, let's separate dev, test, and prod. It feels so wrong.

Thorsten: Yeah. And the funny thing is, it's the same with Spaces, by the way. The feedback, it's wild, it's the whole spectrum. So, with Spaces, I have a screenshot saved on my phone, I haven't shared it yet, so, again, Ian Landsman, he was in my replies, he's like, ‘I don't understand Spaces.’ And the two—yeah,

Quinn: Yeah, we love you, Ian, by the way.

Thorsten: Big fan, Ian. The other two replies right below his were like, ‘I love Spaces, this is exactly what I was looking for.’ And the other one was like, ‘Oh, Spaces are so good.’ And with Boxes, when I released this there were people like, ‘Oh, so is this, you know, do you host your apps? Is this replacing Kubernetes and whatnot?’ And at the same time, while they were saying this, the number of Boxes in our customers' deployments just went up because they immediately got it, and I looked at kind of what they're doing and it's internal, you know, CRM apps, it's internal wikis, it's internal survey apps, it's internal documentation, little app—somebody hosts their personal website on it now. There's so much of this software around that previously you ran on these platforms that, you know, you don't need. And that's—you said this, right? That was your idea. ‘Why can't I turn the Boxes off overnight or something? Like I should be able to turn them off.’ So I added this feature and it's such a cute feature that kind of drives the point home, which is my personal box, it turns off from 11:00 p.m. my night to, you know, 5:00 a.m. the next day. I don't need it. Like I don't need my gym tracking and my notes app and whatnot overnight.

Thorsten: I can save money on it. Sure, like if you have a cloud thing, they scale to zero and then they shut off and blah, but man, it's just a computer, like that stuff works. You pay 18 bucks a month and you can run five apps on it, ten apps, fifteen apps on it, and that does it, and the agent can see it all and help you with all of it, so—yeah.

Quinn: Yeah, and it uses basically bare compute. And that, it turns out, is a lot cheaper than sandbox compute, and the amount of compute, the cost of the compute you need to run the dinky little apps that actually make your life or your team a lot better, they're so cheap to run. So there's someone who actually told me, ‘Well, I actually got a Mac mini, and I could just, you know, run everything on that and save a lot of money.’ But dude, like a Mac mini is like 900 bucks, and you can get a box for like 18 bucks a month. It's going to have to run for a long time until that.

Thorsten: Yeah, yeah. And you turn it off overnight, and then you reduce the cost and it's 10 bucks a month, you know, and it—yeah.

Quinn: So there's another thing, which is, you know, basically, yeah, Boxes are not here to replace other things. We think that there's a new kind of software that you've had the urge to go build, and there's just so much friction, and you don't build it. And now you can build it. Like I made something that's now in the sidebar of Amp. You can put things there, and you can hover over, and you can see the preview, and it shows what's deployed, because we have all these like macOS and Windows and Linux and whatever apps. It shows like when was each one last deployed. I made one that shows what is everyone working on right now. Someone on our team is doing it to like manage our accounting and all of our customers. We're doing our bug report system in this. And what's really cool is they can all talk to each other. They're all built into Amp. You can sign in with Amp to each of them, so then your box application knows who the user is. It could even run inference on their behalf. And it's like the small-scale cloud platform. I hate that framing, but like it has a lot of these primitives that actually are really valuable for you to build the kind of application you need, so that you don't have to go, you know, redo like how does it do inference, how does it identify users, how does it map your team from like your LDAP or your SSO.

Quinn: And we're just getting started on that. I don't think customers have fully explored that yet, and there's some more things we need to ship to make that possible. But that's really cool, and then they can all talk to each other.

Thorsten: Yeah. Yeah.

Thorsten: It's very excited, especially with—just to say this out loud, so we talked about this a bunch before that in Amp, when you report a bug, and if you do this a lot, I urge you, you know, someone who sends bug reports, ask yourself, do you want the Amp team to spend time on this, or is it truly worth somebody's attention that you don't like this button? Anyway, so previously, people could submit bug reports in Amp. We would store them in our Amp application with diagnostic bundles like logs and other information. And then, we would get like a bug identifier. We could hand it to the agent. We would have a plugin so it can talk to our API, and it can fetch the diagnostic bundle and debug stuff. You can already build this yourself, like you can build this for your own app. The problem was that we were at an advantage here because, you know, we were using Amp and modifying Amp, and we could, basically Amp to us as the Amp team is malleable software, like that is the biggest force in this team that if we don't like something we can change it, and ten minutes later it's live. The biggest force. And yet, our customers don't have this force. So, with Boxes, the idea is to lean into this and kind of distribute this force or make it available to others by giving them this malleable software where you have the Orbs on one hand and they can deploy to something on the other hand, and it's just a constant—you see something in a Box app, you have the little widget, you can say, ‘Fix this,’ it spawns an Orb, so it fixes it in production, and you can hit reload and it's live.

Thorsten: Customers should experience this. So, with our bug tracking, what we've decided yesterday is let's rip out our built-in bug tracking and build it on top of a Box. That way we dogfood how it feels like for customers to develop on Boxes and use them to build a real application. And I think the underlying theme is like dogfooding is an incredible superpower and I think Boxes will get better by us building more stuff, production stuff on Boxes, and it's just a great example of Jellyware or instant software or something where it's an internal tool. Any frontier model right now can one-shot a good bug tracker, and it's the tight integration between your agent and the software that makes this really good, you know. That's the whole differentiation that little widget and you close the loop. And I talked with Dario yesterday about the widget. The widget is not about the form of the widget or like feedback or something, the widget is—

Quinn: The widget is a little thing in the bottom right when you're looking at an app or a portal on Amp, and you can click it, you can, you know, send prompts back to the agent.

Thorsten: Yes. Send prompts to, yeah, you can either start a new Orb or now with Boxes you can run a thread on the Box or send a message to a thread on the Box. And Dario released this last week as the feedback widget. So if people have ideas on your website, they can kind of send feedback, or if you have feedback and whatnot. And the point is not like the inputs and outputs, the point is that you want to close the loop, and you don't want to have your app be somewhere over here and how you develop it be over here. I think it's fantastic to use your app and then hit that button and say this paper cut, let's fix it. You want to remove the friction, right? And get it as close as possible, so that's the whole thing. And

Quinn: Yeah, the joyous thing that we can do is, when we're using Amp, which we do all the time, if we see anything, anything that we don't like, then it is so easy for us to go kick off the agent from that, to go and make it better.

Thorsten: Yeah—that's what we need to bring to customers.

Quinn: I think the test for anyone listening would be go and use your app. If you find even the smallest little thing, do you feel joy or friction at the thought of kicking off whatever your process is to go fix it? If it's a Jira ticket, you probably don't joy. If it's an agent and you know that it's going to be fixed in 15 minutes perhaps, you're probably going to feel joy. So, report back. Let us know how that works.

Thorsten: Yeah. It's a lot of fun, and I think there's going to be some people who listen and are like, ‘Oh yeah, I have a workflow for this. I take a screenshot, I file a ticket, and in the ticket, the agent picks it up and then blah and all of this.’

Quinn: Yeah. All right, I have a question.

Thorsten: Yeah.

Quinn: So, we talked about how a lot of people want different things from the Amp sidebar, different ways that Amp works. If Amp can put a little icon on the left side and I can click it, and that runs my app, and that app can fetch from the Amp API, would we ever let people have Amp build a custom sidebar for them or a custom way to look at their Threads if they don't like our UI?

Thorsten: Maybe. You know, who knows? I think it's a—we talked about this, that it's the idea of like software you can now generate with the push of a button, why shouldn't you use this force and have your software benefit from it, you know? Like I think Amp should be—you know, Pi pioneered this, software that changes itself. And I think with Amp, we're seeing this, too, where people want to modify Amp and customize it to their liking, but that's kind of hard, like I don't know, like I think there's a lot of open questions. It's hard to have like malleable software that's deployed as a multi-tenant app to multiple people. And, you know, we've all heard the Jira joke of, well, the true Jira hasn't been tried, you know, the Jira that's set up correctly, you should see my Jira, something like this. So, configuration is not a solution to the problem, but I think like, yeah, maybe like if you're able to build like other pieces of software that can communicate with the main thing, maybe that's the solution. You know, like maybe there's something interesting here where I can build like my custom frontend for Amp, or my custom file view, and have it talk to Amp or something. Like maybe that's, you know, I'm saying a lot of maybes because I don't know, but I think it's really fascinating that what if plugins are not plugins that are loaded at runtime, but what plugins are now is software that talks to your thing and you just build pieces around it, maybe that's the

Quinn: Yeah.

Quinn: And there's this weird tension, because I think that there's a lot of people that would be able to make our sidebar better for their particular use case as an individual, but if you're using Amp on a team, if you're using any agent on a team, probably the person who would be making decisions about how to customize the agent for your company, maybe you don't think that they have good taste.

Thorsten: Solution.

Quinn: And so you'd rather have them be constrained by what we have decided. And we found that with Amp, where we said, you know, for the longest time—now there's some controls, but for the longest time, it was impossible to restrict what models your users could have access to. So inside of all these big banks, these big enterprises using Amp, Amp was the only way that they could get access to like 4.5 Sonnet right when it came out. And people love that, and that was legitimately good for the bank that we took away some of that choice.

Quinn: So there's that principle. There's, you know, the possible—well, does this mean we're going to have a lot more API contracts? But actually it's a really interesting case to think of what is the feed of information, like the automated changelog? Do we need to just ship the raw source code of our sidebar? Well, no, we also need to then ship the human intent that went into the changes that might not be captured in the diff or the commit message, cuz we've found that to be so valuable.

Thorsten: Interesting.

Quinn: And so we're starting to think of how can we enable this mass customization at scale while exploiting the fact that agents could possibly make this thing that three months ago would have been an absolute mess, actually possible and enjoyable.

Thorsten: Yeah. Yeah.

Quinn: It's really interesting. We don't know what to do and some of us are, you know—it's a very tempting idea to just make the Amp frontend open source or something like that, but we would also need to bring the Threads along and we don't want to do all of it for security, you know, whatever reasons.

Quinn: So I don't know, it's an open topic. Maybe a future episode we'll have more.

Thorsten: Yeah, we'll talk about this. We should—I mean, another thing to mention is open source, and I don't think that's the solution, but I don't know.

Quinn: Yeah, and maybe it would be public, you could see it, but not, you know, yeah. So many things are changing.

Thorsten: Just to riff, I think there's also something where what if I'm thinking in weird terms here, but what if you could give your agent or your customers, you give them an MCP server or something like this with which they can inspect your codebase? Or, you know, like something like this where it's like you say to your customer, ‘Guys, we got to migrate from A to B. Use this MCP server, because that way your agent can ask our agent a bunch of questions about the codebase and then they can talk to each other.’ And I think it's going to work. Like I think that would be—it's wild, but it would work.

Thorsten: So.

Quinn: I was just thinking of this. We use WorkOS, and I would like it if the way they do OAuth screens would let us make it so that only users in a certain organization could OAuth. And it seems like the kind of thing they have, and I would love to be able to ask something, is WorkOS working on this, or is there something I missed from their docs, cuz occasionally they can get out of date, that's in their source code? And WorkOS, would they be comfortable with something that would respond to a trusted customer like us, you know, no more than 250 per day, it would all be audited, and each response would be no more than 400 characters?

Quinn: I think that companies could get comfortable with that level of information sharing, and of course like it would be covered under legal agreements, and people are reputable and everything. And that is really interesting. That is something that actually could open up all of these codebases that, frankly, support teams are unable to support, the documentation is downstream of the actual code in the agent. So why are we doing all of these non-sources of truth when we actually could get comfortable with sharing derivations on the fly from the source of truth? Maybe we should do this in Amp. Make it so that you can let,

Thorsten: Our agent answers your questions by looking at our source code, and we put like a diff-based filter around it so it doesn't reveal source code, or something like this. And that way, you know, it's ultimate code search.

Thorsten: But okay, so we talked a lot about like the software's faster produce. It changes a lot of stuff. Let's end on a sad note. Let's talk about what, you know, agents do not solve yet, because we've run into this week, and I think that topic is a good, you know, slice of the period that we're in, or it's a good view, which is we shipped Spaces last week. We also shipped service accounts, I think also last week or the week before. We shipped Boxes this week. We changed our Slack integration last week to work with service accounts. And now we find ourselves

Quinn: We also made it so that in a multiplayer thread, you can say certain MCPs like, you know, our Gmail integration, your teammates cannot use your Gmail if you happen to make a thread multiplayer. And all of these things, each of them, genius idea in isolation, but what happened?

Thorsten: Yeah.

Thorsten: Yeah, what happens is there's conflicting things and there's like no clear answer where, ‘Oh, like somebody created a space and they want to talk to a shared box, but the space was created on, say, Quinn's behalf, but I was the creator of the shared box or something like this, and now Quinn's space, when I write a message, it spawns of—’

Quinn: I'm getting a headache even hearing you say that.

Thorsten: Yeah, whatever. Like basically the problem is there's conflicting permissions or there's conflicting rules, or they're not fully specified out and they're not designed as a whole. And that's bad, and it feels bad, and we need to fix it ASAP. And hopefully, by the time this is out, it will have been fixed. But the point is no agent is going to help you with this. There's no way—like you can ask agents and say like, ‘Look, this is the problem.’ You can dictate for 20 minutes and you can say, ‘Give me a holistic view of all of the things.’ But at the end of the day, we ran into this using agents and we ran into it because we didn't have coordination and we didn't have a plan fleshed out, and no agent said, ‘Stop it. We need like an RBAC system,’ or whatever it is. And it's not a software problem. It's purely a business and specification problem, and it's purely a coordination problem and making trade-offs, you know? And that's still hard.

Quinn: Why do you say no agent will help with this? Is it that no agent will say stop and refuse to add complexity on the shit pile, or do you actually mean that no agent will help with this?

Thorsten: I mean, what I mean is, first of all, they do stop, and I've had this yesterday where I fixed something and the agent would say, ‘Actually, Quinn added this because of, you know...’ And I'm like, ‘Okay. So, let's not do this then.’ So, it does help you navigate this. They can answer all of the questions. They can answer—I had it this morning, a couple hours ago, a user had a question. The agent can go in and say, ‘Well, the user ran into this, but that's the bug because—and our docs are out of date, but we changed this yesterday.’ So, they can help you navigate this. But, when you need to sit down and change the whole system, you really need to think about what you ask of them and how you approach the problem, I think, is one thing. And I would also argue that they kind of led to the problem in the first place, maybe, because it's so easy to make one-off decisions for like five people that then result in this mess. And obviously, the haters, the non-AI-pilled people, will say, ‘Well, that's always the case with agents. They make a mess and they don't know the business logic.’ But I would argue it happens less than you think. But yeah, that stuff is still hard.

Thorsten: It's still hard. And our solution is—well, I complained, as I said.

Quinn: Yeah, how do we handle—like, walk us through, what happens here? What's going to happen now after we sign off,

Thorsten: And—oh, yeah. So, what happened was, over the week, we got bug reports coming in with questions. We got people on X asking us questions. Internally—

Thorsten: Yes.

Quinn: And we also saw these things too, and sometimes because we use so many feature flags, we're not sure if, ‘Oh, is that just because of some feature flag, and don't worry about it?’ And it often is, but you gotta like steel your brain to just be like every single one of these, we need to look into it because if we're hitting it, we want to fix it before a customer gets to it. And like that diligence, that's like one of the hardest psychological things of doing this. But yeah, back to you.

Thorsten: So, we got the bug reports. We got the customer reports via email. We got on X, whatnot. Then we shared these internally like in our bad channel. We were like, ‘I don't understand why I can't do this. I don't understand why I can't do this.’ And it started to pile up. And then we had multiple threads, and then we kind of coordinated to have one thread. And then, I think it was yesterday, I said, ‘Can somebody own this end-to-end?’ Like, I don't want five people, somebody looking at Slack, and somebody looking at Puck, and somebody looking at Spaces, and somebody looking at Boxes. I need somebody to look at the whole thing, including Orbs, including Boxes, including Spaces, and think of the whole thing, and then design something, and then we can implement it. And I said like, ‘I would really appreciate it if somebody does this.’ And then Lewis raised his hand, and he's like, ‘Okay, I'm going to own this end-to-end.’ So now, I guess if we were a large corporation, we'd call him the DRI, but now he's the person who owns this. And that

Quinn: And so Lewis, basically, he has set up a bunch of meetings and is planning on like an RFC process to fix this, right? No, what is he actually doing? What do you think he did, or do you know what he did?

Thorsten: Means—

Thorsten: No.

Thorsten: Yeah. I think what he's doing right now is he's looking at all of the different bug reports. He's looking at all of the different problems. He's writing down and making explicit the things that are the problems. And then he's going to propose—

Quinn: And also, he's got all of the bug reports that people have sent in to help understand what was the customer understanding or expectation or intent because with anything security- or permissions-related, you have to understand what is the customer actually trying to do, and then that maps onto the permission system. But you really want to know that original intent. And we have that. We have people on X and email and in our Slack channels.

Thorsten: There's a lot of information. And I think what's going to happen now is, it's not that you won't touch this, or that I won't touch this, but I think now we will ping him or let him know. Like, I'm not going to go in and change this dramatically, but if I bump into the issue, or if I have to do like a one-off fix again or something, then I will let him know. It's just I'm going to route information to him, I think. And then, hopefully, you know, he comes up with a solution. And then, you know, the solution to a lot of problems is just talk about it, like make it explicit. I keep saying this, like, the first step with a lot of stuff is make it visible. If there's a problem, and you don't know where it is—start making it visible. Start making it so that everybody can see the shape of the problem, and then take the next step. Like if you don't know what to do, start there. And I think now we're kind of doing this, you know? Like we've started to make it visible. Now we have somebody who says like they're going to own it. Let's figure out the next steps.

Quinn: Yeah. And when we put a problem up there, or anyone on the team puts a problem up there, the threshold for someone taking it on is literally just: does it give them energy? And at a lot of other companies, it's like, does it fit on a team's roadmap, and do all of them feel excited about it and all that? And so just with an agent, one person can do a lot, and you want to get into that place where it's so easy to put up these problems. And some people like different things. Like, some people hate doing this work. You would hate doing this kind of work, for example.

Thorsten: I hate it.

Quinn: You accidentally thought that I meant you were going to do it, and you were like, ‘What? I'm the last person to do it.’ But I think he gets energy from it.

Quinn: And there's an interesting thing here where this is going to collide in a pretty rough way with the work that I want to do right after we sign off, to go and make it so that people can be in more than one, or collaborate in more than one workspace at once. I mean, that's a huge complication adder to our security model. And this is one of the hard things. I'm probably just going to go talk to him and let him know, and also see what he's doing, and try not to step on toes.

Thorsten: Please do. Please do.

Quinn: I will not just roll back all of his improvements.

Thorsten: No, I think—I don't know.

Quinn: Oh, by the way, it gets me to another idea. We talked so much about just in this time, how valuable it is to have the record of human intent so that you can go fix things better, it's not just the code change. And when I think about—we don't really get a full picture of what users want. What we see is how that maps to the settings that we have today. And we get bug reports, but I think in the future everyone believes that the way that you're going to work with a computer is not by, you know, clicking around, or at least it's going to get more like that. If we made it so that the primary way to work with Amp's settings, to manipulate things in Amp, was to just talk to it or prompt it, I think that we would get so much richer context about what the user would want to do.

Quinn: And I am tempted, crazy idea, but what if we just entirely kill our settings screen and say the only way to change your settings in Amp is to go talk to Puck?

Thorsten: Yeah. That would be good. Lot of risks here, but there's something—you know, like the risk is, I do think like a UI still provides like some benefits, you know, like some things, but yeah, it would be—

Quinn: Well, yeah. It's like clicking approve or reject when the agent makes a diff. Yeah, so many benefits. You know, I get what you're saying. It's a radical idea, and there's two parts of my mind. Maybe there's a genius idea that we can find that truly just makes this so simple, but maybe the scope of an agent, forget a coding agent, of an agent that touches everything you do at work, that is where your team lives, like think of Slack, but it is ten times more consequential for the work you do. The surface area is so large that it is actually not possible for a human to comprehend all the settings that you need to make.

Quinn: And that's kind of scary. But what I would say is I know a lot of people that configure their computer, that install their computer using an agent now, and that's kind of dismissing the settings screen in their computer. And Amp's scope is the same amount of scope as something like macOS, you could say.

Thorsten: Interesting.

Quinn: And think of what we could do if we capture more of the human intent and I think then that there's ways the agent could better map what the user wants in terms of settings to what we actually can provide.

Thorsten: Yeah, that's interesting.

Quinn: And that's how I feel today when people are doing so much more in Amp in two months than—is it even worth humans worrying about that? Maybe they want to know a few things like, ‘My stuff is secure,’ and that's it. And also, the settings screen, it embarrasses me even though we've tried to make it better. It embarrasses me.

Thorsten: Yeah.

Quinn: I've made Thorsten uncomfortable.

Thorsten: Well, yeah, the settings stuff is weird.

Quinn: It is.

Thorsten: But the other thing is to me, like, what you were describing is like at the end of the day, like, why have the user configure the software if the agent can do it for you, right? And then my counterpoint would be, well, there's value in seeing possible options, or like, seeing things and like discovering things, but then the counterpoint to this is, well, but only if you use it yourself, and if the model does more of this, then—you know, like it's this we have settings for themes, right? Okay, do I want it to ask Puck to change the theme without having seen the themes before? I don't know, like I don't know.

Quinn: Well, what if Puck on-the-fly generative UI to make a little thing where you can click some buttons and then it will update the theme?

Thorsten: Okay, fair enough. Fair enough, but then my counterpoint would be, why even set a theme if you're just talking to Puck? You know what I mean? Like—okay, maybe you have a theme for the chat, but like there's a circular thing going on where the more you remove of the software and put it in the model, the less you have to configure about the software because it's more in the model. You know what I mean?

Quinn: And then it gets to the really uncomfortable question that everyone has, which is, why do you even need an agent to make software if no one else will use software?

Thorsten: Yeah, why do you even do this? Yeah.

Thorsten: Yeah,

Quinn: And yeah, that's where partly—well, yeah, that's why we think that the real economy can benefit so much, and that's where we are spending a lot of our time. And partly that's just, you know, you look up at the stars on a dark night, and you just cannot even comprehend.

Thorsten: Exactly. Exactly.

Thorsten: Yes, that's the counterargument, right? It's been one and a half years since we started this. And it hasn't happened yet. What if it's another one and a half years? And then, what if it's another one and a half years? And then, at some point, you have to say, ‘Well, we've been saying this for five years, and it hasn't happened yet, so.’ I don't know. Who knows?

Thorsten: All right.

Thorsten: Let's end it here. We talked about a bunch of stuff: the future of software, the future of self-hosted software, Spaces, Boxes, complexity, all of this. So what we want to leave the audience with—that I'm going to be in London next week. You know, me and a bunch of other people from the Amp team. So, maybe we'll have time to meet up. Let's see. But then we'll all be in Sydney in December, right?

Quinn: Yeah, week of December 7th. So if you are an Aussie, then come say hi. We'll send out some stuff, and would love to meet folks.

Thorsten: Yes. All right, then we'll maybe see you in Sydney. All right. Bye-bye.