The founder of the company that terrified software engineers is now telling them to relax. Scott Wu, CEO of Cognition, has been making the rounds arguing that AI coding agents shouldn't replace human developers—a curious position for someone whose product was marketed as the world's first AI software engineer.
Wu's apparent volte-face isn't altruism. It's pattern recognition. After eighteen months of deploying Devin across enterprise clients, Cognition has learned what every AI company eventually discovers: the gap between demo and production is where hype goes to die.
The demo problem
When Cognition unveiled Devin in early 2024, the demonstration videos showed an AI that could autonomously complete entire coding projects—debugging, testing, deploying. The implication was clear: junior developers should update their résumés. But Wu now acknowledges that autonomous operation works brilliantly for narrow, well-specified tasks and fails spectacularly when requirements are ambiguous, stakeholders change their minds, or the codebase has accumulated years of undocumented technical debt.
This isn't a Cognition-specific problem. Every AI coding tool—GitHub Copilot, Cursor, Amazon CodeWhisperer—runs into the same wall. The models are exceptional at pattern completion and mediocre at understanding why a particular architectural decision was made three years ago by an engineer who has since left the company.
The collaboration thesis
Wu's revised pitch centers on what he calls "cognitive partnership"—AI agents handling the mechanical aspects of coding while humans focus on design decisions, stakeholder communication, and the judgment calls that require understanding business context. It's a less revolutionary vision than "fire your engineering team," but it has the advantage of being achievable.
The economics support this framing. Enterprise clients aren't looking to eliminate developer headcount; they're looking to ship faster with the headcount they have. A tool that makes existing engineers more productive is an easier sale than one that promises to replace them but requires months of supervision to work reliably.
The credibility question
Wu's restraint is notable precisely because it's unusual. Most AI founders are locked in an escalation spiral, each promising capabilities more transformative than the last. Anthropic talks about systems that could automate entire job categories. OpenAI gestures toward artificial general intelligence. Against this backdrop, a founder saying "actually, humans remain essential" reads as either refreshing honesty or strategic repositioning after the product underperformed.
Probably both. Cognition reportedly raised substantial funding at a valuation that assumed Devin would capture significant market share from human developers. If the product is now being repositioned as an augmentation tool rather than a replacement, the company's growth narrative needs adjustment.
Our take
Wu is right, even if he arrived at the conclusion for commercial rather than philosophical reasons. The most durable AI applications are the ones that enhance human judgment rather than attempt to substitute for it. Software development, like most knowledge work, is less about the mechanical act of writing code than about navigating organizational complexity, managing competing priorities, and making decisions under uncertainty. AI handles the first part well. The rest still requires someone who can attend the meeting where the product manager changes the requirements for the fourth time.
