Cognition

In late 2023, Scott Wu and his cofounders—Stephen (early engineer at Scale AI) and Walden (early engineer at Cursor)—began exploring the next frontier of AI capabilities. The founding team, mostly reconnected through competitive programming circles, shared a conviction: reinforcement learning would be the paradigm shift that moved AI from pure imitation learning (like ChatGPT's initial training) to systems that could actually *do work* and learn from feedback. For Scott, who had previously spent five years running Lunch Club (an AI professional networking product), the moment felt ripe.

They believed code was the perfect domain because of one crucial property: automated feedback loops. When you run code, you instantly know if it works or fails—the ultimate RL signal. "Code has this whole automated feedback loop, right? Where you can run the code and that is the kind of automated feedback that really feeds into the RL," Scott explained. Starting in November 2023, they booked an Airbnb "like a hackathon" and began hacking.

The journey was anything but linear. "We've pivoted like eight times or something," Scott said, laughing. Initially they tackled contest programming problems, building an agentic system to solve coding puzzles. But the real insight came at the second hacker house: a full software engineering agent that could integrate with the tools engineers actually use—GitHub, Slack, Linear.

They launched in March 2024 after roughly four months of non-stop work. At launch, Devin was roughly "a high school CS student." The team quickly realized that raw capability improvements were only half the battle; the other half was designing *how humans interact with an autonomous agent*. Early versions simply took a task and returned finished code. That felt impersonal and disconnected. Scott and team realized they needed to treat Devin as a collaborator with personality and presence.

The breakthrough was the name itself and framing. "Devin as a name really is our attempt to kind of capture the soul of that as a product—treating it like a bit more of an autonomous entity that you can hand off tasks to, that you can work with, that you should be teaching and learning with over time." By December 2023, they made Devin fully available self-serve, and in recent weeks rolled out Devin 2.0.

The article doesn't detail the initial customer acquisition story directly, but Scott mentions that they "work with companies of all stages and sizes"—from one-to-two-person startups building prototypes to Fortune 100 companies and public banks. The product's ability to integrate natively with GitHub, Slack, and Linear (the tools engineers already use) likely made adoption feel natural for early customers. The framing of Devin as a "junior buddy" rather than an alien tool also resonated.

Early assumptions about how engineers would use agents proved naive. Scott noted that people would see a blank screen and be paralyzed, unsure how to hand off their first task. The team had to lean into pedagogy: "Get folks to think more like, whoa, whoa, whoa, let's work on getting the repository set up first. Let's make sure we hand Devin a couple one-pointer tasks so it can get familiar with the code base."

What worked was the asynchronous workflow. Initially, teams watched Devin action-by-action, but the real productivity came when engineers kicked off multiple Devins in parallel and only jumped in to steer course. On Cognition's own team, each of the 15 engineers works with roughly five concurrent Devins, which collectively merge "several hundred pull requests into production in the Devin code bases every month." About 25% of all PRs are now Devin-authored, with the team expecting this to exceed 50% by year-end.

Other product wins included Slack integration (immediate, familiar notification), GitHub integration (seamless PR workflow), and the "Devin Wiki"—an auto-generated codebase understanding that both Devin and humans could reference. Interactive planning phases and the ability to touch up code mid-task also proved essential.

Cognition is using its own product to build its product at scale. Devin has evolved from "high school student" to "college intern" to "junior engineer" in a year. The capability jumps were dramatic—reasoning improved, tool use expanded, error recovery deepened—but so did the product experience.

Scott is bullish on the future. He rejects the fear that AI engineers will replace human engineers, instead invoking Jevons Paradox: as engineering becomes faster, the total demand for engineering explodes. More personalized software, more products, more possibilities. "I think there's going to be way more programmers and way more engineers a few years from now than there are today."

He sees software engineering evolving from "brick layer" to "architect." The mechanical work—debugging Kubernetes errors, migrating code, fixing bugs—will increasingly be handled by AI. Humans will focus on defining problems, architecting solutions, and specifying intent. Programming fundamentals remain essential: understanding databases, garbage collection, TCP/IP, abstraction layers. But the *form* of programming will transform. "Pretty quickly, the form factor of what it means to be a programmer obviously is going to change."

Cognition itself is the living proof. With just 15 engineers shipping Devin 2.0 and managing hundreds of async tasks, they're operating like a company 10x their size—a glimpse of how engineering teams everywhere may soon work.