Scott Wu of Cognition AI on Devin becoming teams' top code committer and the coming golden age of software engineering

Summary

Scott Wu, CEO of Cognition AI, makes the case that software engineers are entering their most powerful era yet — not being replaced by it.

Cognition launched Devin publicly about a year ago, spent much of that time working through the messy realities of enterprise codebases — logging systems, documentation, tooling, migrations — before opening general availability in December with a self-serve credit card plan. The early results are striking: some teams now report Devin as the number one code committer in their repository.

The original bets

Cognition was built on two founding convictions. The first was that reinforcement learning would work — a contrarian view at the end of 2023, when most AI products were still built on imitation learning, essentially predicting the next token from scraped internet text. Wu draws the contrast sharply: RL is closer to AlphaGo-style self-play, where the model runs a problem hundreds of times and learns from right and wrong, rather than just mimicking. The second bet was that the product paradigm would shift from Q&A and text completion to genuine agents — systems that don't just answer questions but interact with real tools, run code, read logs, and iterate the way a human engineer would.

How Devin fits into engineering workflows

Wu's framing is that engineers today spend roughly 10% of their time on the interesting work — deciding what to build, reasoning about architecture, solving hard problems — and 90% on Kubernetes configs, unit test failures, dependency upgrades, and repetitive migrations. Devin is designed to absorb that 90%, leaving engineers to operate as architects rather than bricklayers.

The workflow integration is intentional: Devin reads Datadog logs, sits in Slack, and gets tagged like a teammate. Crucially, it builds institutional knowledge across sessions. If one engineer asks Devin to handle a version upgrade, it can draw on a similar migration it completed the previous week for a different part of the same codebase. Wu sees this compounding codebase context as a meaningful long-term moat — the gap between a brilliant new hire on day one and a ten-year veteran who wrote half the code themselves.

On the broader competitive landscape

Wu is diplomatic on DeepSeek and the pace of model competition, but his practical view is pointed: if a team isn't using AI coding tools today, it is already behind. He expects AI adoption to follow the pattern of cloud and mobile — a hype phase followed by deep structural integration — with code as one of the first verticals where that second phase is clearly already underway.

On the vibe-coding platforms like Lovable and Bolt, Wu draws a self-driving analogy. Those tools represent cruise control: useful for a wide set of tasks, but you still need a driver for everything else. The fully autonomous layer — where you describe what you want in plain English and step back entirely — is coming, but complex enterprise codebases still require engineers who can direct the system with precision.

The bigger thesis: single-use software

Wu's most forward-looking argument is about software abundance. Today it makes no economic sense to commission an engineering team to write a script you'll run once. Within a few years, Wu expects that to be routine — you describe the task to Devin, it builds the tool, runs it, and discards it. Niche products serving a hundred users, or even just one, could be built to the same quality standard as YouTube or Instagram. The constraint on software stops being production cost and becomes imagination.

For a 22-year-old entering the field, Wu's advice is to go deeper into computer science fundamentals, not away from them. Understanding abstraction layers, breaking down problems logically, reasoning about systems — that judgment becomes more valuable as the mechanical work gets automated away. The theory of CS, he argues, matters more in the agent era than it did before.