Cognition demos Devin, billed as the first AI software engineer

Cognition, a start-up founded by Scott Wu, introduced Devin, an autonomous coding agent it called “the first AI software engineer,” able to plan and execute multi-step programming tasks with a shell, code editor and browser inside its own sandboxed environment. The company said Devin could learn unfamiliar frameworks by reading their documentation, build and deploy full applications, find and fix bugs in existing open-source codebases without supervision, and complete real jobs posted on the freelance platform Upwork.

The claim that anchored the launch was a benchmark result: on SWE-bench, an evaluation built from real, unresolved GitHub issues drawn from open-source Python repositories, Cognition said Devin resolved 13.86% of issues completely unassisted, against roughly 2% for the previous best-performing system. It reported an even larger gap when competing tools were handed the specific files that needed editing — a task Devin was not given that advantage on — where rivals still managed under 5%. Cognition released demonstration videos showing Devin working end-to-end on tasks including training and fine-tuning a language model and contributing a fix to the mathematics library SymPy.

The demonstration drew rapid, largely uncritical coverage treating Devin as evidence that autonomous software engineering had arrived. That reading did not hold up under later scrutiny: independent reviewers who examined the released footage in detail, most prominently the YouTube channel Internet of Bugs, argued in 2024 that several showcased tasks had been misrepresented — including one in which Devin appeared to have been given a solved problem to “fix,” and another where a claimed success concealed an underlying failure — concluding the demo materially overstated what the system had actually accomplished.

Whatever the gap between demonstration and reality, Devin’s launch marked a shift in how coding assistance was pitched: from autocomplete and chat-based suggestion toward an agent expected to complete tasks independently over an extended session. Cognition built on that positioning in the following year, culminating in its 2025 acquisition of the coding-tool start-up Windsurf, after Google paid to license Windsurf’s technology and hire its chief executive.