Hi, I'm Scott. I'm here from Cognition. We make Devin, the AI software engineer. I wanted to talk a little bit about what we call the golden age of software engineering. So, a little bit about us. We're Cognition. We got started about two and a half years ago.
A lot of what we do, obviously, is we make Devin, the AI software engineer. We work with a lot of the biggest companies around the world. And so, as you can see, you know, Goldman Sachs, Mercedes-Benz, Lowe's, Elevance, etc. We work in all the different geographies. We're big in the US, in Europe, Latin America, Asia Pacific, Middle East, and so forth.
And so, you know, the natural kind of first question I would ask is, what do all of these folks have in common? You know, what do Goldman Sachs and Mercedes-Benz and Lowe's have in common? Well, the simple answer is, in 2026, all of these companies are software companies because every company is a software company, right? In the same way that we don't say that companies are internet companies anymore, we don't really say that companies are software companies anymore. All of these folks have tens of thousands of software engineers. They have tons and tons of things that they're building. You know, for Mercedes it might be all the little details that go into the car, that can go into the software, that can go into the process management, right? For Goldman Sachs, it might be all the transactions that they process or, you know, all of the internal tools that they're building and using. But all of these have tons and tons of software, right?
And so, the first question that I always ask folks in terms of how to think about this is, look, what would you do with an infinite army of software engineers? And I would ask all of you guys to spend a second and just think about that. What would you do with an infinite army of software engineers?
And I ask this, and then people kind of assume that, okay, I'm asking like a theoretical question, and yeah, it's cool to imagine in a few years like what's going to happen. I think everyone's kind of seen that as a whole, you know, AI coding has taken off. I don't know if people have understood really deeply just how much it has.
And for us internally, one of the craziest stats, which is one of my favorite stats to share, is that at Devin, first of all, we use a lot of Devin to go build Devin, unsurprisingly. But our own amount of software that we've shipped has literally exploded. It's gone crazy in the last six to eight months. And so you can see over the course of the six months from November to end of April, we literally 10x'd the amount of software that we shipped, okay? We brought in some more engineers, you know, you can see we added about 40% to our engineering headcount, which was meaningful. We went from like, you know, 45 to 60 or something like that. But obviously the much bigger impact is in what software's done. And remember guys, software engineering and AI coding was already a big thing in November 2025, right? So, what changed exactly over the course of six months that allowed us to ship 10 times more and do 10 times more? Well, I think a few things in particular. First of all, the models and the core capabilities just keep getting better. And so I think the single biggest thing that I always think of when I think about this is this METR study, which looks at, for an agent or for an AI, what is the average length of a task that the AI can do without any interruption, right? So, it can do it all in one shot, it can figure out all the details. And as you can see, not that long ago, a few years ago, what we were talking about was, 'Okay, it can do about 30 seconds of work.' You know, if you take a look at GPT-3.5 right in the middle of that chart, 30 seconds of work, which is not bad, by the way, right? And that was the ChatGPT moment, that was the exciting thing. It's just you ask it a conversation, you know, you have some questions, and it can have that with you, right? Now, we actually kind of broke the scale. You know, it turns out the longest tasks in the data set, they don't go that much beyond 16-hour tasks, but the latest models kind of just go and solve all this, right? And it's worth thinking about because obviously when you're talking about 30 seconds at a time, you are talking about an assistant. You're talking about a chatbot, right? You're talking about something that you can ask a question, it'll give you that answer in 30 seconds. And you're talking about something that can do 16 hours of work or even multiple days of work is kind of what we're getting into now, and then you really are talking about a full AI software engineer, right? You're giving it an entire task, you're giving it a project, you're asking it to own things end-to-end and go and do that.
And it's not just any one model, which is worth calling out, it's actually all the different models. And so you can see, you know, this is each of the different labs and their flagship models and how they're each doing. Every single one of them has obviously gotten much, much better. And so we're getting to a point where there's real saturation on a lot of tasks. If frankly, all the models are good enough to just go and do these tasks, right?
And what does that mean? Well, for us, what that means and how you really get the meaningful gains here is by having the agents close the loop. Okay? And so previously, one of the big things that folks were talking about was writing the code, right? And the first big breakthrough I would say in AI coding was GitHub Copilot, it was tab complete. You were looking at the files, you're having these things go through. Now, rather than just writing the code, you are doing every single step, right? Autonomously. What that means obviously is you can have the agents take on this entire thing. And so it's you're tagging the agent on a ticket or there's a customer that had an issue and you're just asking the agent to go solve that issue, right? You ask Devin to go solve that issue. And it's going to do every single one of these. It's going to understand the codebase. It's going to plan out the change. It's going to write the code. It's going to have computer use. So it's going to go be able to test the code by itself. It's going to be able to do review, right? And that's how you get the real gains.
And you know, with that too, it means you can have them initiate their own work, right? So you can start their sessions autonomously. You can have these things go end-to-end. One of my favorite stats. So obviously, how did we 10x our software development rate? We have Devin write it all. And so as you can see, you know, you see the percent of code committed by Devin going up to 90 plus percent. We basically use Devin for pretty much everything right now. The thing that's even crazier is Devin actually also does the majority of starting Devins.
So, this is all of the Devin sessions at Cognition by week, and this is what fraction of them were started interactively, in other words, by a human that said, 'Okay, hey Devin, can you please go do this?' And then this is what fraction of them were started asynchronously because there was an automated event, there was a trigger, right? And so, every single thing you can imagine, it could be Devin starting sub-agents, it could be planning out specific tasks, it could be, for example, like if you have, you know, API routes that you want to monitor for performance, then yeah, you run automated triggers to go run those routes, check how performance compares to what it was one week ago. If it's somehow noticeably slower, then you spin up a Devin to go take a look, right? If you have an error in Datadog, you spin up a Devin to go take a look, right? You run all these automatically. And that's what we're seeing as well, where the majority of the sessions can now actually be run and kicked off proactively, too.
So, all of that is to kind of say, okay, agents can act like real software engineers now. My favorite story on this, by the way, is, I was in Japan recently to get to speak to some of our Devin users there. Someone asked me this question, we were doing a kind of like Q&A session. Someone asked me this question, which was, 'Hey, so, what honorific do you think I should be using with Devin?' I said, 'Well, what do you mean, what honor?' And she said, 'Well, like I just talk to Devin all the time, I just don't know, should I say Devin-san? You know, should I say Devin-kun? Is it Devin-sama? What's the right one?' But you can really think of your agents as a full software engineer, right? They can use the computer themselves, they can work with every single integration you have, they can go and like test the code, they can do browser use, they can do all these parts, right? What does that mean? Well, obviously the next question then is, how should you reframe your org with that in mind?
And a few things in particular. First of all, as we're saying, you want to do things as parallel and as event-driven as possible. And so, this is a lot of what we've seen has really taken off in the last three to six months, which is all of the kinds of things, you know, it's almost like every kind of mental hook that would make you as a software engineer want to go and do something, you can set up that hook for Devin as well, right? And so, for example, you know, bulk CVE remediation, right? All of us have tons of tools that go and scan through the codebase and help us figure out, okay, what are potential warnings, what are potential security vulnerabilities? You can have Devin go and automatically fix those, right? Big version upgrades or big codebase refactors that you're going and doing at scale, you can have Devin go and tackle every single file of a massive codebase, right? And you kind of run all these in parallel.
Same thing for all of these event-driven cases, right? Bug triage. Customer report comes in. Usually what happens is, okay, a human has to go take that and say, okay, here's the thing, let me go start off a Devin to go do that. Well, you can actually skip that middle step and say, okay, look, just have that go straight to a Devin and then you can have a pull request ready for you that you can go and review and go and do. And then similarly for all the everyday development that you do, you know, working on your PRs or handling testing or just kind of building out a particular feature end-to-end, you can do every single one of these obviously autonomously, right? It's worth calling out because it's the, you know, if I were to just pull this up, the meaningful changes that's happened in the last six months, I would say, is, you know, by the way, 13% in December of 2025, 13% is a lot, right? Every single team, you know, think about a team of a thousand software engineers and imagine you could tell them, okay, you are getting an extra 130 software engineers worth of bandwidth to go and build things. You can do a lot with that, right? But that means taking on kind of like a lot of the one-off features or some of these particular repetitive tasks. I think the important shift that's happened is actually now you can have that work for pretty much everything, right? And 89% or 90% is actually not like an extra 130 engineers. I mean, if anything, it's like you are doing 10 times as much, right? Because you can have, you know, for a thousand human engineers, you basically have 9,000 AI engineers that go along with that.
And so, when you get to that point, what you also, of course, care about is you want to think about, okay, well, how do I evaluate these based on real business impact, right? And so, you know, I think the whole token maxing era was...
lasted for approximately six months in my mind. And obviously people are talking about, okay, how many tokens are you using? How many... It's almost exactly the same as saying, okay, how many lines of code did you write? You know, there's a famous story, right? Which is, you know, back in like the '90s or early 2000s, there was a software engineer and all their teams were measured on how much code that they wrote. And so, at the end of each week they would kind of submit a report, okay, here's how many lines of code I wrote. And then one of the engineers submitted a report and said -200. They said -200. So, then they asked, okay, what does that mean? Well, I actually just re-implemented this whole thing way faster, way cheaper, it was a way better architecture, and actually I was able to go delete like 800 lines of code for the 600 lines of code that I wrote. And then they said, okay, you know what? Lines of code maybe not the right metric. But what we're seeing is every single one of these is moving towards, obviously, what actually matters for your business. And this is kind of why we come back to that core question of, okay, well, what do you do with this infinite army of software engineers, right? What is the additional software that you build? And I think it's worth kind of like really intuiting this because I think one of the most common questions that we hear, as I'm sure you all can imagine, is, okay, well, so what happens to all the human software engineers? Well, the simple answer is, look, every human software engineer is 10 times more productive, but guess what? We have 10 times more things to go build, you know? And when we think about all the things that we want to go do, it's, you know, imagine going to the Department of Motor Vehicles, except it works, you know? All of that is software, all of that is process, all of that is things that can be better. Imagine going and logging in and pulling up your medical records, except the product experience is actually good, you know? Imagine logging into your bank and actually being able to do all the things that you want to go do, right? And so, every single company out there, obviously, has these teams of thousands or tens of thousands of software engineers because there is so much software to build. And the reality is there's actually 10 times more of that to go do, right? And so, a lot of it is holding agents to those same standards of, okay, well, how do I think about the actual, you know, project outcomes that I'm going and delivering? How do I think about what I'm doing end to end? How do I think about basically what it is that I want to go build and why I can do much more of that?
And so, you know, I think we're kind of already moving past this token maxing era, which lasted all of, you know, January to May or so. There's some of the famous kind of stories or folks, you know, Uber, for example, very famous is said they burned through their entire token budget for the year in just three months. Obviously, I think the answer to this is not to say, okay, well, your goal is to do as much with tokens as possible. The answer, of course, is to figure out, okay, well, what are the outcomes that you want to go and drive, right? And so, I think a lot of what we're getting to is what we're seeing is you are thinking much more about what are the outcomes that you're seeing? And what is the productivity? What are you doing with the capacity? Why does that matter, right?
Which brings me to the end of this, which is, look, the golden age of software engineering is really here. And I've had folks ask me about this all the time, okay, you know, all this stuff with AI coding is cool, but when do we really get to abundance? Guys, we're getting there right now. And if your team is going and doing these things correctly, you literally can do 10 times more, right? And I think the core question, I think, for all of us to ask and for all of us to think about really is, what do you want to build? So, thank you, guys.
