Cognition CEO Scott Wu says Devin, the company's AI coding agent, now commits 89% of the code shipped by Cognition's own engineers — and he still rejects the framing that the product is built to replace programmers. Wu's two-year-old startup raised $1 billion at a $26 billion valuation this week, on the strength of a pitch that Devin 'naturally owns tasks end to end.' The remaining sliver of Cognition's code comes from local agents inside Windsurf, the rival coding tool Cognition acquired last year.
The same blog post announcing the round described a shift to 'a world of self-driving software development,' a phrase that lands awkwardly against Wu's softer public messaging. Asked whether Devin could stand in for an L4 mid-level programmer, Wu told TechCrunch the answer was yes and no — capability is there for some tasks, but framing the product as a human replacement is not the company's pitch.
Wu places Devin's current skill level 'somewhere between a junior and a mid-level engineer,' depending on the task. That capability band is what makes the 89% internal-commit number meaningful: Cognition is eating its own dog food at scale, and the dog food is doing most of the cooking.
Key facts
- 01Cognition raised $1B at a $26B valuation, just two years after the startup was founded.
- 02Devin now commits 89% of the code shipped by Cognition's own engineers; the rest comes from local agents in Windsurf.
- 03Cognition acquired AI coding rival Windsurf last year and folded its local-agent tech into the Devin workflow.
- 04Wu pegs Devin's capability at 'somewhere between a junior and a mid-level engineer' depending on the task.
- 05Cognition's blog framed the round as a shift to 'a world of self-driving software development.'
Wu's framing is that Devin is a tool for engineers, not a swap for them. He keeps a small stuffed-animal Devin mascot on his desk and described the agent as 'my buddy that helps you build more.' He started coding at nine, won a nationwide math competition as a second-grader, and came up alongside other technical prodigies including Scale AI founder Alexandr Wang — context Cognition leans on when it argues the company is run by engineers building for engineers.
The practical sales pitch is more granular than the marketing copy suggests. Wu says Devin's natural lane is long-tail maintenance work most engineers actively dislike: bringing legacy systems up to date, porting applications between platforms, the toil that fills sprint boards but rarely makes a resume. Freeing humans from that work, he argues, pushes them toward 'much more of the creation side.'
That's the bull case for AI coding agents as augmentation rather than substitution, and it sits uneasily next to the broader 2026 backdrop. Multiple tech CEOs have used AI capability as the stated reason for layoff rounds this year, and Cognition's own 89% figure is exactly the kind of number a CFO at a software company will be asked to explain on the next earnings call. Wu's position is that Cognition's intent has never been workforce reduction; the intent of the buyer often is.
The recursive question — agents that improve themselves and climb the seniority ladder autonomously — is the one Wu sidesteps most carefully. He calls the trajectory 'a wild ride' and extends the pattern beyond software into customer service and medicine, while insisting the human should remain the decision-maker.
Cognition's $26 billion valuation puts it in rare company for a two-year-old startup, and the round arrives during an active stretch for AI coding tools. Anthropic just shipped Claude Opus 4.8 with a dynamic-workflow feature aimed squarely at agentic coding, which we covered earlier this week. The competitive frame is no longer 'can an agent write code' — that's settled — but how reliably an agent can own a multi-step task without supervision, and at what price.
The skeptic's read is that 89% of internal commits at a 200-person startup building its own coding agent is not the same workload as 89% of commits at a Fortune 500 bank with a 30-year-old codebase, compliance constraints, and a security review process. Cognition has not published a third-party benchmark to accompany the funding announcement, and Wu's own description of Devin's skill ceiling — junior to mid-level — implies a meaningful share of senior engineering work remains out of scope.
For the AI coding market, the Cognition round resets the price of admission. A $26 billion valuation for a two-year-old agent company, on the back of a Windsurf acquisition and an internal-deployment proof point, raises the bar for what a credible agentic-coding play looks like to late-stage investors. Wu can argue Devin is a buddy rather than a substitute, but the number that travels is 89% — and that's the number every enterprise buyer will quote back when they ask their own engineering org what's still being done by humans, and why.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
