Were you expecting that level of attention? And from your perspective, what was that moment like?
Yeah. Well, it is, by the way, you're doing just great as a software engineer. And I obviously think that will continue. I mean, we're all engineers ourselves. And, you know, we're certainly excited for... I actually think we're getting into really the golden era of software engineering, but you know, it was crazy. I mean, obviously, we were not even close. I mean, we hoped the launch would go well.
If the implementation layer is fully taken away, what does that mean for software engineering as a career?
And I mean, I think the answer is actually absolutely yes. Um, and I think the reason that is, is because, you know, for as long as it's us using the computers rather than the other way around. And hopefully it's that way for some time. You know, it's going to be up to us to decide what to do, right? And I think actually at the, you know, we call it computer science, we call it programming, we call it software engineering, but at the core, the thing that it's really about is about just being able to tell your computer what you want it to do.
Welcome to the Why You Should Join podcast, where every two weeks we sit down with the founders of some of the most notable hypergrowth startups. As always, my name is Asan. I am one of the co-authors of the Why You Should Join newsletter and I'm also the host of this podcast. Today's company, Cognition, is building the world's first AI software engineer. Their flagship product, Devon, is an AI agent designed to autonomously carry out complex software development tasks end to end. The company was founded in November of 2023. They've raised over $300 million to date with investors including Founders Fund, Kosal Ventures, El Gil, Conviction, and more. Not to mention, their most recent fundraise brought the company's valuation to a whopping $4 billion. I'm super excited to be chatting with Scott Wu, the company's CEO and co-founder. Scott, thanks for doing this.
Yeah, thank you so much for having me on.
Absolutely. Well, let's dive right into it. So, I know you guys had a demo that did quite the rounds. So, I suspect just about everyone here has seen what you guys are building, but just in case to set the stage for those who are familiar with Devon, haven't used it or seen it. Do you mind giving a quick run through of what exactly it is that you've built?
Yeah. Yeah, for sure. So, we're building Devon. It's a fully autonomous AI software engineer. And it's funny saying this now in 2025 because I think when we were saying this in 2023, everyone thought we were kind of crazy, but yeah, I mean the whole idea of Devon is a full asynchronous coding agent that can take on full tasks, right? And I think there are a lot of different products out there in code today. The thing that really sets us apart as an AI coding product is that Devon is able to execute on the full end to end of the stack, right? And so Devon will not only write code, edit code, and so on, but you know, it'll be able to run the lint itself, it'll run the tests, it'll execute commands, it'll install packages, it'll look at tickets, it'll talk to you in Slack, and basically use all the same tools that a human software engineer would use.
Wow, that's awesome. And we're definitely going to dive into a lot of that very shortly. But while we're on the topic of sort of like the demo video and all that, I was curious honestly just personally, I think that demo video and that launch was a little bit of a cultural moment. I think like your guys' launch really sort of broke the chasm between tech Silicon Valley bubble and even had my mom texting me saying, 'See, I told you you should have been a doctor.' Um, so I mean obviously when you release something the goal is for it to sort of pop off, but I'm curious, were you expecting that level of attention? And from your perspective, what was that moment like?
Yeah. Well, it is, by the way, you're doing just great as a software engineer and I obviously think that will continue. I mean, we're all engineers ourselves and you know, we're certainly excited for... I actually think we're getting into really the golden era of software engineering, but you know, it was crazy. I mean, obviously, we were not even close. I mean, we hoped the launch would go well. You know, we were putting that internally honestly for us it was such a mess. Uh, and it was one of my favorite weekends. Um, one of my favorite weekends ever, honestly. We the article went out on a Tuesday morning. It went out at 9:00 a.m. The thing is it was actually supposed to go out at 6:00 a.m. Um, it was like a Bloomberg article and everything, but then we had to... I asked them like we wrote them last like please, please, please, could you just like hold we need like a couple more hours? And you know, we did need those hours to get to 9:00 a.m. And it was like, um, you know, there's the whole the video of me talking at the couch and everything and then showing what Devon looks like and stuff. And like, dude, we hadn't finished filming that video by like 1:00 a.m. that night. So, there was like... and that's, you know, there's obviously all of the other, you know, there's all the other content and sharing the demo runs and putting out the blog post and stuff like that. Um, and it was... there was so much to do and at the time it was all very much just like, you know, using every hour as best as we could. I mean we were a very small team that is still a very small team. Uh, and a lot of it was like as you can... I mean there's just so many things to do, you know, with any launch right? There's getting our investors or our friends to help out and promote it for us. There's, you know, putting together the blog post and really articulating like what we think the future is going to look like. There's getting it out to the initial early users so that they can try it and give their own firsthand feedback of what the product was like. And uh, and there was just that... Yeah. Yeah. It was a fun couple days for us. It was like, I mean we were honestly shocked to see it.
No, and a little bit of a follow-up on that too, which is yes, the initial shock must have been pretty insane, but also I think with it came a lot of sort of lingering hype and obviously hype I think is great in some ways. It can tangibly help with, you know, raising capital or hiring or other things, but I think also brings a little bit of a target on your back. It invites a lot of the skeptics, the critics, and you know, stuff like that. How has that been sort of from your experience kind of navigating all of the media onslaught that came afterwards?
Yeah, for sure. I mean, it's obviously, you know, once you put it out, you're not really in control of your own narrative anymore, right? And it's how it is. And you know, we felt this way this entire time that, for example, that this was going to be amazing for software engineers and the people were going to be able to build way more and do way more. But we were pretty quickly accused, I would say, of basically, you know, of trying to take software engineers' jobs or all these other sort of things. You know, there's the fear, there was the doubt, the skeptic, you know, there's a lot of these things. Um, and I think for us the, you know, we obviously do everything we can to just kind of like build a really great product and show the world what it looks like, but I think the biggest access that we have control over is just getting the product out there and letting people try it out for themselves. Um, and so the big thing is to kind of just push more and more towards that. And so, you know, in December, just a few months ago, we had our full general availability launch where, um, you know, anyone could put down a card and just sign up and pay for Devon and connect their GitHub repo and get started, right? Um, and before then, you know, it was primarily enterprises. Um, but yeah, I mean, it was really exciting basically to just get it out there and have everyone see all the full aspects, you know, the good and the bad and everything and just try it out. And I mean I think it's very... look, I mean I think Devon is certainly not magic, you know, and it does some things very well, some things not so well and so on, but I think all in all just letting people try it out and see for themselves what the experience is like was the big thing. Um, and in the last couple months since then it's been fun to see all these different, you know, a lot of different companies where it's like Devon is now the number one GitHub contributor at the company, you know, them posting the GitHub graphs and stuff like that. And I think it's been great basically to have it just be much more organic that way.
Yeah. Yeah. And for folks who maybe are not familiar, Devon is generally available so people should definitely check it out. Um, and you know you can do so on their website. But on the point of sort of controlling the narrative, um, maybe let's dive into that a little bit like sort of what is the big picture vision here. You know, what is endgame for Devon? Is it that just about anyone, no technical background, can sort of come in and take an idea and get a working application in a matter of a few prompts? Or is it a little bit more like, hey, you're an engineer, you do have technical background, but now you sort of have this agentic assistant or maybe like an army of these interns who you can kind of guide and do work for you.
Yeah, it's much more so the latter today. You know, it's very focused at a technical audience and it's, you know, practical engineering teams with their existing code bases almost always, right? that are, you know, figuring out building new things, fixing issues, you know, doing upgrades, whatever it is on their codebase. Um, and working through that. Uh, the way that we kind of think about it is, you know, I think self-driving is a great analogy here where um it's, you know, there's been a lot of different phases of self-driving, right, of self-driving cars, right? And, you know, it started with automatic shift and then there was cruise control and so on, right? Um, and I think eventually in the long run, you know, we're now reaching this point, I'd say, with self-driving where there comes a point where you actually don't have to have a driver's license at all. You know, you can sit in the back seat and just have it do everything. Um, but up until then, it's obviously still really valuable to know how to drive, right? Um, and I think software engineering is kind of like that, but even there's even a lot more dimensions, right? Where it's kind of like, yeah, you know, you might know a lot about front end, but less so about back end. You might know, right? And so what we typically see with Devon is, yeah, it's folks who are pretty technical who are using Devon themselves. And it's more like turning the bricklayers into architects, if that makes sense. You know, it's more like turning every engineer into an edge lead and giving you your own team of Devons that you get to work with and assign tasks to and you know, they'll ask you questions or they might need guidance from time to time, but it's basically having your own army that you can work with.
And so that's definitely what it looks like today. I mean if you play out that scenario further and you know everyone is sort of the architect, you know at some point if the implementation layer is fully taken away, what does that mean for software engineering as a career? Or even asked as like an education, like if you go to school, you know I went and studied software, what do I study in school now?
Yeah, people ask me this all the time, you know, so should I even really still be studying CS? And I mean I think the answer is actually absolutely yes. Um, and I think the reason that is, is because, you know, for as long as it's users rather than the other way around and hopefully it's that way for some time, you know, it's going to be up to us to decide what to do, right? And I think actually at the, you know, we call it computer science, we call it programming, we call it software engineering, but at the core the thing that it's really about is about just being able to tell your computer what you want it to do, right? Um, and I think what you learn in computer science for example in CS classes in school, sure there's a little bit of understanding this detail, you know, these programming languages or things like that, but by the way, programming languages change all the time. You know, it was Python. It's Python today. It used to be, you know, there's Java, there's C and there was, you know, assembly and basic stuff before that too, right? Uh, and I think in practice like the core skill I think that really remains I'd say is just this base kind of problem solving, right? It's like how do you break down problems and understand things? How do you take a problem and understand okay here's exactly what the problem is and here's what my solution is going to be and here's how I'm going to build it, you know, how do you understand the model of what's possible to implement and what architectures work and don't work and so on.
And I kind of think what we're going to get to is a point where using Python feels kind of like using assembly today. You know, I mean there are some people who do use assembly, right? Especially if you're really relying on clock cycles or whatever it is, you want to peel back the layers of abstraction and really work with the kind of the base gadgets, right? But I think at a high level, you know, I think what happens kind of today is that high level understanding of these layers of abstraction of this is what the computer is doing under the hood, but I'm able to talk with it in a much easier language is what most programmers are using. And if anything, by the way, there's way more programmers than there were in the age of assembly. You know, people actually don't realize this, but up until around the year 2000, there's actually still fewer than a million software engineers, and today there's about 30 million in the world.
Yeah. And so it's grown quite a bit, right? And obviously that's despite or perhaps because of programming getting so much easier, right? And I think it's going to be the same thing, you know, I think there will be a point where you don't have to look at the code if you don't want to and a lot of it is just like interacting with your own product and saying, okay, you know, we should have a new page here that does this and that and like by the way, let's track all this information in this form and you know what would be really cool is like all right, let's collect this data and then have it show it to the user this way and that way and whatever it is, right? And it's just happening basically, you know, in front of you because you're working with Devon to do that. Um, I think it'll take us some time to get to that future. But I also think that the core problem solving skill and the core thing, you know, I think we'll still call it programming actually, you know, because it will still be this core piece of building software and interactively being a problem solver. So yeah, that's really interesting. Well, we'll definitely touch on sort of like what it means to be a problem solver and how that kind of shapes up in like sort of the advancements in reasoning models as well, but kind of on this point of, you know, you'll still, you know, software just being abstracted and you'll still be able to use this almost like a no code or AI coding tool. There was this interesting point that I think Bill Gates made, I forget where it might have been an interview or something like that, where he was sort of talking about how, you know, a lot of these no code AI coding tools, automation, anything that kind of abstracts away the technical implementation from the developer mostly helps with familiar problems and they don't necessarily push software development forward in the same way that say maybe a new programming paradigm or major infrastructure shift does, which is like hey, you know, these tools know what already exists and kind of piece that together for you. Do you think Devon sort of fits into that which is like, hey, this is a final form of like implementing as we know today, but or do you think Devon can also kind of drive, you know, these breakthroughs that previous no code or automation tools couldn't?
Yeah. And I mean, Bill is the goat obviously, so he knows a lot more than I do, but what I would say is um yeah, you know, I think that's the way I guess I kind of think about it is it's true, but it's always been a spectrum. And that's kind of how I think about it, right? It's I think we've kind of always said this that the, you know, the human is the one driving and there's still kind of like a ton, you know, there's execution parts that the computer takes care of, right? And you know, I think that's essentially how things have... that's how humans have always worked with computers, right? But at the same time, you know, what counts as ideation or what counts as core problem solving or what counts as execution, I mean that's changed a lot, you know. I mean it used to be the case for just to give a simple example, um, you know, totally unrelated to AI. It used to be the case that if you wanted to have your own website then, you know, you had to go host a server, you had to go figure out things with DNS, you had to go like, you know, set up things, there was no Cloudflare, there was no... I mean all of these various things, right? And obviously as time has gone on, we've been able to abstract more and more of it away and get closer and closer to the real meat of just being able to decide what it is that you really want, right? And I think that I think the same is the case here, you know. I think that um I mean in general in the field for example like I think AI has shown its ability to solve some incredibly difficult problems, right? And you know you could call it pattern matching but it is I think there's some really difficult and really involved stuff, right? But I think it's maybe one way to put it is even if all it did was, you know, taking care of the 90% of the execution and so then you got to do the 10% of ideation 10 times more, I mean that would be enormous. I also think that, you know, as time goes on we will refine our own opinion on what it means to be idea versus execution, right? And we'll get to the point where, you know, within that 10% a lot of those things we would think of as kind of like, you know, layers that you can decide or that you could abstract away or things like that, right? And we'll continue to peel that back and we'll go from 90 to 99 and 99 to 99.9. But I think that um but yeah, you know, I think humans have always kind of worked with computers in this way and yet it is very different, you know, using a computer today or writing software today than it was 20 or 30 years ago. So yeah. Yeah. And so I guess kind of the point that you're making here is that look hey, you know, no matter how much we abstract there will still always be the human that's using the computer and sort of what that results in maybe isn't so easily predictable right now but the human will be in the driving seat is like safe to say.
Yeah. Exactly. I think folks sometimes say something along the lines of like oh, you know, there's, you know, computers they can't do things that humans can't do or things. I mean, you know, my computer can do a lot of things that I could do... I can't do, you know, I can't multiply these big numbers or run through all the... and so you know I think practically the divide has generally... it's always been kind of like a bit more fuzzy of a divide and I think at a high level it is something like, you know, the human is the one directing and coming up with the ideas and the computer is the one doing the execution. But yeah, you know, I think the border of where what we count as ideas and execution, I think that changes all the time. So yes. Yes, that's a great point. Yeah. And I guess on this kind of introduces like a little bit of like the reasoning models and there even I know this is a bit of an unfair question, but playing a little bit of like even like a devil's advocate on this, right? Which is that just say these models get great, right? Which is and to the point where what do you think that is there a world that could exist where the human really doesn't have to do much beyond the initial idea? Even I know we're talking about like this border of execution versus idea, but you know even like edge cases or how to improve it or how to optimize. What if the reasoning model could cover all of that, right? Which is you really just have to have the idea and the rest is made.
Yeah. No, I mean it's going to be an interesting few years. You know, what I would say is it is I think we're going to... we're about to see in the next two or three years, you know, the level of magnitude of change that probably took place over the last 40 years in software where yeah, I think a lot of these things, you know, we're going to basically reinvent entire new frameworks and new processes for how we build these things. Um, I think that um I think that we'll yeah there will be a lot of spots where it's going to be an interesting question basically of what is possible with AI and what is not possible with AI and um you know so far we certainly haven't seen any signs of slowing down which is uh it's crazy that the things are still going as fast as they are um at the level things are at. I think there's a real core of being a creator um which I think is it is what makes humans, you know, really unique I would say uh and that is the thing, you know, it might be the case and by the way, you know, this is the case with technology throughout time obviously, you know, we have very different needs today as a society than we did a thousand years ago, right? Um, but I think that uh yeah, you know, it might be the case for example that a lot of these basic things are now provided for us in the future because this technology takes care of itself in certain ways and so on. But you know, I think the innate kind of human passion and the desire to create and to do things and, you know, there will still be a lot of things that are scarce and there will still be things that have meaning and um and I think that uh I think that'll be more similar than people sometimes portray it as basically.
Well, I guess maybe it's safe to say that Cognition and Devon is not the end of software engineering, maybe even the start of something even huge. So that is pretty exciting. Yeah.
No, I mean I agree and I think it's uh you know again it just comes back when I think about, you know, all the products out there that could exist and one way that I kind of think about it is, you know, there's a tier of products which billions of people today use, you know, I'd be talking about like YouTube, Instagram, you know, TikTok or things like that and these are incredibly well-made products. I mean to be on it's like, you know, they're streaming you gigabytes of data, their algorithm is really on point, they never go down. Every single piece of the UX is as intuitive as possible. You know, it's super super engineered to be, you know, a really great experience, right? And those are really really well-made products, right? And then, you know, it's kind of like every tier, every order of magnitude you go down, you kind of see the difference in product quality, honestly, right? You go to the products that have like 100 million users out there, right? And then you go to the products that have 10 million users, then 1 million users, and then 100, you know, and then eventually it's like, you know, your local... it's like your community like school website or something and it's, you know, it has the address, it has a nice photo of the page, but it's certainly not as polished of a product experience by any means, right? And I think that uh to me it just kind of speaks to how much more there is out there to do with software, right? I mean, I think there are so many other great products to build where it's obviously I mean it just doesn't make sense to build right now because it would take months of software engineering time to do that, right? When and you know you can imagine a point where a lot of software is actually built single use, like that is something I think we'll see in the next few years which I think is going to be really exciting where it's, you know, software that exists for only for a single purpose and then you use it that one time and then, you know, it's done. But I think in general the, you know, people talk about Jevons paradox all the time. I think it is there's nowhere that it's more true than in software. You know, we have always found more and more and more ways that we want to use software in the world.
Absolutely. Yeah. That's awesome. Well, to drill a little bit, I wanted to dive into a little bit of like a more thought experiment kind of side of the podcast. And this is a little bit for our more research-inclined audience members. Um, I want to touch on the topic of like maybe even reasoning models. Um, and I have sort of two kind of not very related questions um to this, but sort of the first one was I thought you had a really interesting conversation with I believe was the CTO of Together AI and you guys kind of touched on the topic of specialized models. Um, I'm curious like do you see a future where Devon isn't just one general AI software engineer but instead maybe a collection of these specialized models optimized for its own kind of side of software, you know, development? Um, and if so, would Cognition need to sort of build its own reasoning model to make this happen?
Yeah, so Devon is already it's a compound AI system, which means that it relies on multiple different models that all work together basically. Um, but yeah, you know, what I'd say in general with reasoning is it's worth reflecting upon still that RL works like it is insane actually. I think the, you know, we call this all the last few years the Gen AI era, right? I think there are actually two eras, you know, there's one between 2022 and, you know, let's say six to 12 months ago, and then there's the one that we're in now, and they're actually quite different eras. You know, I think the previous era was much more about imitation learning, right? It's basically you train a model, you read the entire internet, and then you get something that can talk like the internet basically, right? Which is a big deal. Well, I mean it was obviously ChatGPT went crazy and all of the products, you know, for those couple years. But I think now we have a totally different paradigm which I think is a lot closer to how humans think about things, right? Which is reinforcement learning basically which is like if you have a very clearly defined set of tasks and you have a clean feedback loop for how you decide if the task was successful or unsuccessful, you can basically have your model try that task over and over and over and try the other tasks and do them all until it's learned how to do all of them very well, right? Um, and it's a very different paradigm. And if you think about what's next, well, you know, um some folks would say, you know, that's already AGI or ASI or whatever you want to call it. I think it's to some extent it actually kind of is. Um, but I think the main thing that kind of is next obviously is what do you teach it, right? What is the benchmark that you train it against? What are the environments that it learns on? Um, and a lot of I think what we're going to see over the next little while actually is basically this new paradigm where these models are specialized for the tasks that they're trained on, right? If you have a very clean clear loop of here are the things that I want this to be really good at and let's say you're building AI for customer service, you're building AI for sales, you're building AI for legal or you know whatever space it is, right? If you have like a really clean loop, you know, we're pretty rapidly getting to the point where in fact you can just train a model that just gets like 100% on that loop, right? Um, and then the thing that happens obviously is that in some sense the differentiation is actually, you know, what you define as the behaviors and what are the set of things that it uses to make decisions and what tools does it have access to, right? Um, and so I think for us it's how we've always thought about these things because I think it's um I guess one way to put it is, you know, I think the model training side and the actual product behavior side are just so tightly intertwined where, you know, it's a very different thing to teach someone to be a really great engineer than it is to teach someone to be a really great lawyer and so that requires a specific training loop and it requires specific data points to learn on and so on. So yeah. Yeah. And Okay. Okay. So, well, I said these questions were really related, but I think maybe now they are a little bit more than I even initially assumed, which is, you know, to me this almost kind of connects to the point of like sort of verifiable reasoning, which is like, okay, if you have Devon and it's capable of, you know, debugging, iterating, learning from his mistakes, you know, one could see it almost be a tool for generating like high-quality like ground truth data or something for a reasoning model, right? Do you think that's possible just like a Devon's usage could generate sort of high-quality training data that maybe could be leveraged to build better reasoning models?
Yeah. Yeah. Yeah. For sure. So, one of the nice things about code and this is by the way I would say, you know, kind of the main reason that code I think has been the first to really take off in terms of language model use cases. One of the great things about code is you get to run the code, right? And so, you know, with all these other fields where you might need a human to go review the output or, you know, you are solving AI for accounting and your feedback loop is, you know, you find out a year later that you got audited and that was a bad outcome. You know, you're in jail. Yeah. Um, whereas, you know, in code it's basically I mean you run the code and it gives you the result or it doesn't give you the result, right? And it's obviously that alone is not a perfect thing, but it gives you a much cleaner feedback loop um that ties in with everything else in order to make these models learn much faster, right? Um, and so I think there is like a real dip here. You know, even the way that we do all of our emails, for example. Um, you know, we shared some posts on basically how we do this and how we run these emails and so on, but a lot of it is very much, you know, it's okay, here's a repo and you have to get the repo working and, you know, doing that might involve all sorts of various things that, you know, we as engineers know very well, which, you know, you try to run the repo and it turns out you're running the wrong package. You know, it's installed like the wrong version or something. You have to go downgrade the version and then you have this other port open for this other thing. So, you know, and then you have to go fix this little bug and stuff and then you run it. But when you get the result, you know, the result is...
Some text or some outcome or something that is not fakeable, right? And that's the kind of thing that gives you a very tight feedback loop on determining whether or not Devon succeeded on a task. And so that's how that most closely ties in, I would say.
Yeah, absolutely. And I'm glad you sort of made a point of this because I'm sure to you and your team this is very obvious, but I think for a lot of people maybe they don't initially sort of think of how that feedback loop allows this specific domain to really skyrocket. And it makes me personally very bullish on things like Cognition, which is that you have that feedback loop for data, which is very powerful.
I guess now sort of pivoting or segueing into another topic. One of the advice I got was that to be a good podcast host, you kind of have to ask a question that's on everyone's mind, but maybe no one really wants to ask. And this is where we kind of enter into that little bit more tricky area of, let's say, the competitive landscape, right?
I'm curious how you view all of this, which is obviously there's a good number of teams, some big, some small, kind of working on this AI for code problem. Without having you reveal any trade secrets or anything like that, what do you think is fundamentally different about Cognition's approach? What makes you confident that one, Devon is ahead right now, and sort of stays ahead?
I mean, I think the first thing I would say, which is just kind of a high-level commentary on the space. First of all, I think there's a lot of really great teams out there and a lot of really great products. But one thing I would say is code is actually just really hard, and I mean this in both the positive and the negative. But the way I would describe it is the ceiling of code is incredibly high. I think one way to say it is, if you imagine a use case where people are doing a Q&A and it's like, 'Hey, who was the third president of the United States?' And it's like, okay, it was Thomas Jefferson. At some point, there's not a better answer that you can give. I mean, it was Thomas Jefferson, right? Whereas I think with code, we kind of meme about the 10x engineer or things like that when we talk about humans, but it's true, right? There is always another level of understanding that you can have. There are harder things or more efficient things that you can build. There are better ways to build things. And I think it's one of the nice things about our space because it just means that there's so much surface here that there is real differentiation, I think, in terms of the kind of value that you could provide. There is always another level to push and another tier to get to.
In terms of how we think about things, look, I mean, I think honestly we're big fans of most of the other products in there. I think Devon today is still rather unique in that it's this full kind of autonomous agent, right? And we've always had this view on what we think the future of software engineering looks like. And it's that, yeah, I mean, as the capabilities get better and better and better, that you're going to want to do more and more tasks of this form. But I think the way that we think about what we really spike on as a team, it's, you know, we kind of invented the category, honestly. I mean, we believed in agents before anyone else did. You know, there's all of this mess, as you can imagine, to figuring out these questions of how does your agent learn over time? How does your agent interact with humans? How does it insert into the workflow? How does it make the right decisions about the codebase? How does it know how to use tools like the browser or Datadog or CI or whatever in order to make smarter decisions? All of these different questions. I mean, it's a very messy problem space, but it is the one space that we spend all of our time thinking about, you know, and it's what we really care about.
