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

Summary

Scott Wu, CEO of Cognition, says Devin is growing 30% week over week — a figure he attributes partly to the broader agent education wave. As every major player launches coding agents, more users are learning to think asynchronously about AI work, which is pulling demand toward Cognition's core product.

Wu's market framework is simpler than most: there are two markets, not four. Engineers and engineering teams on one side; non-engineers on the other. The consumer-versus-enterprise axis matters less than that split, because engineers behave the same way whether they're working on personal projects or shipping pull requests at work. The "app builder" category targets non-engineers who want to create software without coding knowledge. The engineering market, which is Cognition's primary focus, is more connected and more continuous than the segmentation suggests.

Two types of product work

Cognition draws an internal line between what Wu calls "street performer Devin" — the version of Devin doing impressive demos that go viral — and the core product, which is solving real issues inside real production codebases. Wu says both matter, but the ratio is heavily weighted toward the latter. The demo moment where someone one-shots a calendar app is useful for awareness; it is not the business.

On the broader question of when non-technical executives can instantiate full software products without an engineer in the loop, Wu says both top-down and bottom-up paths will converge. Large companies with existing software products face mounting pressure to close the quality gap — Wu cites flight booking apps and X as examples of how low that bar remains. Meanwhile, individual builders will gain progressively more leverage. The underlying driver is the doubling of autonomous coding session length every three months, a rate Wu describes as roughly 16x per year. That compounding is what eventually makes the "build me Google Reader" prompt realistic.

Competitive positioning

Google Jules, OpenAI Codex, and GitHub's agent expansions don't alarm Wu. His view is that parallel asynchronous coding agents have been Cognition's core bet for the past year and a half, and broad validation from incumbents confirms the direction rather than threatening it. He expects multiple players to persist long-term, but argues coding agents are unusual in that both raw capability and product interface matter equally. Image generation is mostly a capability race. Customer support is mostly a product and workflow problem. Coding agents require both — model intelligence is critical, but so is the interface that lets engineers specify details, check in on work, and review outputs.

The real limiting factor

On what constrains coding agents most — context window size, reasoning depth, or environment management — Wu is direct: managing complex environments is the binding constraint. For contained, single-file problems, AI is already close to the ceiling of what's needed. Where agents slip is when they need outside information: test results, linter outputs, deploy logs, issue-tracking context. That's why Cognition's integrations with Linear and Slack are not peripheral features. An agent that can't access the same systems a human engineer uses is, as Wu puts it, "super handicapped."

He's skeptical that long context windows alone solve the problem. Humans have small working memory but excel at intelligent retrieval — pulling in the right information at the right moment rather than holding everything simultaneously. The next frontier for coding agents isn't a bigger context window; it's learning to retrieve and reason over external information intelligently rather than just ingesting more tokens at once.