Scott Wu on Cognition's 30% week-over-week growth and how the AI coding agent market is splitting into two

talking. Scott is in the studio. Let's bring him in. [Music] Welcome, Scott. There he is. How's he going? What's happening? Good to have you back. Yeah, good to be back. And by the way, I love Dave Tenra. It's uh it's it's it's so fun to be on backtoback with him.

I love the I I love the story of of of recruiting and truly just fighting for the best people.

It's I' I've always felt the same way, which is, you know, people always ask about evaluating talent, which is of course, you know, it's it's a very deep thing, but but if anything, I think the much more valuable thing is just you kind of already know who the the really great people are.

And the question is like how how can you how can you build your whole whole organization around, you know, making it the right one for them to join. Yeah. Yeah. The um sort of a random story, but but uh was reminded of this.

I think we talked about it a couple days ago, but Steve Jobs apparently called the founder, the original founder of Siri like over 20 times. He called him every single day. He knew that he wanted Siri to be a part of Apple and he just repeatedly called over and over and over and over and over and over.

Wouldn't stop until he finally said yes to uh joining. So, but I think some something about it too is like I think to you you said something which is like everybody knows who the best people are. But I actually think that's maybe only the case if you're also ultra No, it's easy. Hiring is easy.

You just look up who's won an IMO gold medal. You hire all of them and then uh you just raise a bunch of money and then the rest is easy. It's super easy. No, it's it's fair. It's fair. Yeah, you have to be you have to be in the right kind of circles and communities to know the best people.

But but I I generally think that people spend relatively too much time I think on evaluating you know people who are at the border uh versus like really just fighting for the the absolute top tier people. So yeah. Yeah. It's always it's always a mistake to think like oh this is like a B player.

I I can make them into an A player. Like there's just so much sweat there and uh the math rarely works out. Usually you're better just finding someone who's a great fit for your company, your team, the skills that they have, and you can just uh turn them loose and let them run, which is always the best.

Anyway, uh I mean, tons of news. Uh every company seems to have an AI coding agent. I want to know how should I be thinking about the market. I was uh I was uh on X talking to some people. I had kind of mapped it out as there is consumer code that is written when I just prompt 03.

So I the example I gave was I wanted to know how tall a desk was in an image. I uploaded the photo and 03 without me asking it to write code wound up write writing a ton of uh image processing code in Python. I didn't say write code. It just wrote code for me. That's kind of autonomous AI coding.

Now there's codeex uh which feels more proumer. I have a GitHub repo. I wanted to go fix a bug for me. Then there's something more in the enterprise side like hands on the keyboard all day long. I need a superpowered IDE. That's Windins surf cursor.

And then I'm seeing the top down enterprise deals from companies like Cognition. So how is that the reasonable frame for thinking about the market or are there is it all just winner take all? Is there one market? Are there four? How are you thinking about it and what are you seeing in the market these days?

Yeah, for sure. So, so high level breakdown. Look, I think there's a few different product experiences. And by the way, it's fun to see so many agents coming out this week. Uh, but there's a few different product experiences, but I think there's primarily two markets is is how we kind of think about it.

And it's, you know, in one line, it's basically engineers and engineering teams versus non-engineers. Uh, I I think the thing about it with DevTool, you know, totally agree with your point obviously that there's there's some differences between enterprise and self-s serve and so on.

Um in practice I mean you know we have both businesses for example like and a lot of these other players that you mentioned also have both businesses. In practice, I think what happens with developer tooling is, you know, engineers are engineers, right?

And so, uh, engineers are talking to the same people, you know, the same people who are building their own projects at home are also going and, you know, shipping, uh, uh, pull requests and stuff at work. And and there's kind of a lot of overlap in this that space.

And so there's kind of the the the I'll call it the like kind of like the app builder category of like you know build me a simple app and and that's kind of like targeted towards you know non-engineers today and giving them the ability to to write software build software um like you're kind of saving and then there's there's engineers which there are kind of some subcategories within but but I I I think for the most part it's all it's all connected much more so my current AGI test I've moved the I've moved the goalpost 25 times I'm moving them again my current and my current AGI uh goal is if I can go to Google's new uh jewels AI agent and if I can with one prompt say build me Google reader and it does it then AGI has arrived that's my new goal to yeah it's kind of insane if you think about so so one of the things you know there there was uh some research that came out and kind of people did some statistical analysis and if you just look at you know the length of time that an autonomous coding system can go uh in between every human intervention It's It's kind of insane.

It's doubled like every 3 months. Yeah. And so it's like that's like 16x a year, you know, and we've been doing this for a couple years now. So we're like, you know, I mean, not that long ago was like a couple seconds at a time and now you're doing hours of work, right?

But but it's like I just mention that because, you know, Google readers obviously it's like a really great product. You have to imagine they put like a million hours into that, right? And there will be a point at which it is, you know, that window a million hours of work. But but yeah, we we'll get there.

We'll get there. Yeah. Yeah. Exactly. Yeah. We had Will Brown on the show and he was saying that uh we we we have AGI but I call it 15inute AGI and so yeah you can go and do something but if it's if it's if it's a task that would take a human a week it's not going to be oneshotable in the in the traditional sense.

How do you and the team think about the balance between wanting to create product experiences that somebody could just be like almost like a demo, right?

Like somebody was posting yesterday that they like oneshotted a calendar app and like that's like cool and and you know exciting and all that but maybe not uh massively valuable.

But I'm curious, do you guys have some sort of like internal framework for how you decide whether something is like really worth spending time on? Is it like a truly important problem uh or or is it just is is an opportunity just an opportunity to go viral and get attention but doesn't create that enduring value. Yeah.

Yeah. For sure. So I mean it's in the long term our target has always been you know let's let's do real engineering work that that is you know that that that teams are actually products that teams are really building out there in the real world.

Um and I think the thing that's kind of interesting is is yeah because it is all you know one group of people. I I think it is uh the ability to to build all these like cool things out does matter. It's uh we have a distinction internally where we call it street performer Devon.

Basically it's the the version of Devon, you know that that that where where people are asking Devon to just do really cool stuff, you know.

Um and it's something that we spend some time on, but obviously the the the core by far is really just like how do we have Devon solve real issues and real projects um for your codebase. Yeah. Yeah. street performer. I like that. I do wonder.

Uh so yeah, it feels like there's still two frames of mind uh this consumer and enterprise even for stuff that I would call it I would call work.

There's certain times when like I don't want to open an IDE or I don't have time in my day as a live streamer and someone who does most of their work over email and phone um more of like the executive role. Uh I I still will want an engineer to intermediate the products that I want to build.

And I'm wondering if there's a timeline for when that stops happening or I start instantiating whole software products.

This idea of like the it feels like most of the time when you hear someone say, "Oh, I used AI and I vibe coded an app in the weekend that they're still a programmer and they're still running a business and they're still working on that 40 hours a week or something like that, but they're just getting a lot more leverage to this.

I I feel like the really cool moment that's maybe coming is when I I I actually say I want Google Reader and it just builds it for me.

Um and so so in terms of shifting that enterprise experience of like real not street performance but real software into the consumer uh into the consumer space does that come top down or does that come bottom up from uh like the chat box that I'm already interacting with? Yeah.

And so so I I I think both forms are going to happen in the sense that you know big companies are going to look look there's there's obviously always competitive pressures.

There's a lot of relative things going on and and I think you know big companies that already have like you know software products that millions of people use. There's going to be a lot more onus on them to just make really really great software. You know it's I mean it's I I I keep track of this every day.

You know when when I'm doing things and I run into just bugs in software in the stuff that I'm using. It's it's it still happens all the time. I mean, let's let's be honest. Try and book a flight. X is uh X is the the absolute worst culprit here.

The amount I I interact with more bugs on X than any other piece of software I use. I completely disagree. I I mean, it's so much worse if you go and try and book a flight on the American Airlines app. The bar is so low. I don't know. I Yeah, go ahead. Go ahead. No, go. No, I was going to say, yeah.

So, so that's, you know, their like product quality is going to have to get way way better there.

I think the other side is also going to be true which is individuals are going to be be able to build cooler and cooler apps but but yeah I mean I think the high level kind of general trend is is um as we're saying you know as these tools get more and more autonomous like you want to be able to just use them for longer and longer stretches of time which is I mean which is which is pretty topical given all the agent launches that we've seen this past week.

Yeah. Uh, talk to me about positioning against Google's offering. My my experience from as a consumer. I only interact with uh, Google right now on the AI front.

As a consumer, uh, obviously I've like run ads on Google from a business perspective, but in general, I see a Gemini launch and then as a consumer, it's very very hard for me to access that product that I have five different Google accounts. I can never tell which one is on the Google AI premium tier.

I got I finally got in. I got access to VO3. It was an incredible model. Feel like it felt like they were truly uh a step above and there was kind of a chat GPT style moment for for video that given that the audio comes through. It's super photorealistic.

It's a great experience and it's and it it justified me not only paying $500 a month, but going through the UX hurdle.

Is is the developer experience on the enterprise side dealing with Google similar or are they more dialed in there because it's more important because I imagine that one of the advantages that you have as a startup in the enterprise sales space is that you can do founder le sales you can take the needs of the enterprise into account much more uh different SLAs's more custom solutions there's so many more things where you can iterate even if Google is able to come out with something that uh is like wow okay they're really good at at research but on the product side maybe it's lacking.

Yeah. Yeah. So, so it's you know Google has jewels now open AAI has codeex launch you know GitHub is doing doing more with agents as well and can I just tell you I mean it's the we have always had the view that just like everybody's going to go do this eventually you know.

No it'd be really bearish if everyone was like oh no Scott's like just kind of crazy. I don't know what he's up to. I don't think it's a good idea. It's like no you should inspire a bunch of people to be like that's a good idea. we're going to do it too and then you should still beat them. Yep. Yep. Yeah. Yeah.

And it's like parallel asynchronous coding agents, you know, has been it's been the core bet of our entire company basically for the last year and a half. Um and I think it's it's really interesting to see. Yeah. And so so I think there are a few factors going on.

Like I I do think there will be multiple players, you know, in the long term. But but to your point, John, I mean, I think there's also just a lot of uh there's a lot of meaningful differentiation.

And I think the thing that's so interesting about the coding agent space is like it is truly you can kind of look at most spaces and and kind of call it I'll say like this is a capability space this is a product space right and to just give you a kind of a few examples I would say you know it's like um image generation is like sure you know there's a lot of kind of like icing you can put on and things like that but at the end of the day you know if you just have a way better image generation model you know that is the that is the the kind of the the core like that is the meat of what you're building right um and like on other spaces for example you know like customer support or something like that you know it's no one would disagree with you that the capabilities are there and a lot of it is just like what is the right interface for humans and AI to really work together and uh and get that going with coding agents it is truly both where it is you know on the one hand obviously yeah I mean every every IQ point really matters you know in the system that you build and every every piece of data that it can use and analyze and work with it really matters um and then on the other hand it is you know it's the software is just messy enough and deep enough that it's it is it's it you do want to have a lot more control of what's going on.

You want to be able to specify these details. You want to be able to check in and review the work yourself and all these other things.

And so there's a lot of deep work with interface and it's uh yeah I mean in short I think these are these are the kinds of problems that we've been thinking about for the last uh for the last year you know and so it's uh it's fun to see and and you know the fun thing too actually is like um I mean we're we're up 30% week over week actually with just the and it's and I think there's kind of a funny thing which is like you know people people know how to use chat bots right uh and it's like something that we've all kind of learned over the last couple years.

I I think that the the agent form factor is just really different, right? And and it's something that people have to learn and have to understand and and now that everybody's talking about agents, it's actually, you know, it's actually been great for us. So, do you think deep research is part of that?

Like I feel like deep research for a lot of people is their first first uh introduction to agentic uh AI in the sense that it's like you fire it off and you let it go for 20 minutes and then you come back, right? Yeah. Yeah. Yeah. For sure.

I think the whole right asynchronous kind of action is is really like a big piece of it and I think it's you know we were talking about this last time but it's like agents are going to be everywhere you know pretty pretty quickly and there's no reason you shouldn't have an agent who's you know going and buying your flight so you don't have to go on the website yourself or going and you know all of these different things.

Um and and I think for a few reasons code has kind of been the first place where that's really taken off but but but yeah this is a total wild card so feel free to pass if you haven't dug into it but uh Ben Thompson was talking about the agentic web. we're talking about agents now.

Uh, and he was kind of noodling through the idea that stable coins could be a relevant tool in the toolkit for the agentic web as you, you know, ask the one of the examples was plan a birthday and it's doing a lot of things on online. Maybe it's buying uh birthday cards.

It's also looking up a cake recipe and then printing that out. Well, that cake recipe website is monetizing with advertising right now. stable coins in the future, you could just pay a tenth of a cent for that page visit. Are you thinking about stable coins at all? Does that sound real?

Just if you nude on it, how do you think about this stuff? Yeah, for sure. So, I mean, I think with, you know, building software engineering age, we're probably, you know, on the further end of away from it, but but with that crypto pivot. Yeah.

Not a lot of Yeah, I imagine that not a lot of customers are paying onetenth of a cent for Devon right now. that's probably more in the said, you know, it is like an interesting thing which is like maybe one way to put it is like a lot of the norms of the internet to change, right?

You know, it's we have captures for example is kind of how you prevent bots and and it's it's almost like captures aren't the right like system where you actually now want your agents to be able to go and like navigate on the internet and do things for you.

And like a simple example of that for us is like this is again totally unrelated to software engineering, but it's like you know we give Devon our ramp card and then Devon just goes and orders groceries for us and stuff like that.

It's like, yeah, it's great, but but you can imagine, you know, there's a lot of these systems of like, yeah, how how should agents interact with money in the real world?

How should they u, you know, h how how products and and and websites and stuff have the right incentives if it's humans that are going to them anymore, but agents and humans? And so, yeah, a lot a lot of questions. Has Deon ever missed a memo on a ramp uh receipt or is Devon pretty good about that?

I think Devon's got it all down. Yeah, Deon takes screenshots and stuff as well, so it's all Yeah. How are you thinking about the the sort of stack, right? Because I know you got a partnership with Linear. Obviously, we work with Linear.

Uh there's idea that uh much like a member of like a team, an agent can kind of appear in a lot different places, right? Uh a teammate can be in the office, they can be in a, you know, in a Zoom call, they can be in Slack, whatever.

Um, and how are you thinking about kind of like a aentic kind of like orchestration, right? And sort of managing agents. Yeah. Yeah. Linear is, by the way, they just had a launch this week. They had launched this morning actually. So, so yeah. Yeah. Shout out to the linear team.

Um, and yeah, it's it's uh it's an interesting one. I mean, speaking of this kind of like product interface question, right?

Um you know the the work that you have to do is really like u is is you know I think one of the difference between chatbots and agents is an agent is truly like it truly needs to be like basically an employee in the same way you know it is really like you as an engineer are using your team of of of junior buddies who are going to go and like implement this stuff for you right um and um and and and so like a lot of what that means is it has to be in all of these same systems, you know, in order to work really cleanly.

Um, and like the the the standard workflows that we have. I mean, we have Slack, for example, we have uh linear and JR and so on. But it is a big part of it. Yeah. Last question uh actually from your post. You asked for questions from people who follow you and Remy Olsen asked great one.

What do you see as the primary limiting factor for coding agents between context window reasoning which I imagine is just the amount of of inference uh you're delivering to reasoning tokens managing complex environments or other things.

Yeah, I think of those I would say I I would say managing complex environments is probably the biggest one in the sense that look if if you like speaking of 15minute AGI you know if if you give it like a a very contained coding problem where it's just like all right write one you know one file of code that solves this thing it's honestly it's already insanely good you know AI is is already basically there I would say on that I I think that the areas where you see these agents or or models in general like really slip up is is kind of stuff where like you actually just need outside information to go and do what you're doing.

You know, it's it's like the simple examples are, you know, if you're fixing bugs or you're implementing things like you want to be able to test the code yourself. This is a big part of why connecting with all these systems is so important.

And so, you know, plugging into Slack or Linear, for example, makes it really easy for you to kick off your dev and runs or whatever it is.

But I think a big part of it too is like if you guys are actually doing your issue tracking in linear or if you guys are you know this is how you do front-end deploys and this is how you run the llinter or something like that.

It is obviously you know the the agent is going to be super handicapped if it doesn't have access to the same things. Um and you know that the next step beyond that obviously is not accessing those things but but learning how to really intelligently work with these and use them to make decisions.

Any reaction to the llama 4 behemoth? I I I think I saw 10 million token context window seem bigger than ever at the same time kind of going back and forth on how that project is progressing. There's been kind of mixed reactions online. What what's been your take away from that?

Yeah, there's a lot of cool projects working on longer context windows.

I I would just kind of give as as one point which is uh it's interesting you know if you think about human uh human human thought which is which is not always the best parallel but but is is is obviously is is the one example that we do have you know humans actually have terribly small context switches you know it's like when you get like the six-digit code you know it's like like I don't know it's like pretty hard to remember what the six digits were and type those in goldfish mode yeah even with coding it's like it's like once you have to remember like the three different files that it takes you a while to spin up the context window and then you don't want to get pulled out of coding.

You want to be in the flow, right? And so the thing that humans are like really good at I would say is more like intelligent retrieve, right?

To your point, which is like you put this file and you're looking at it and you're like, yeah, like I remember why we did this two months ago and then you're kind of like you can kind of like replay it in your head almost. That's kind of the thing that's that's that's really interesting.

I I don't think it's so so I I guess long story short is like I I think long context is going to help a ton. Uh it's going to make things a lot better.

But but I think there are some kind of like fundamental things of like how do you access all of the information that you need access to to make this decision that's right in front of you right now which is like context alone doesn't solve that problem. Sure. Well, thank you so much for coming on. This is