Cognition AI

Contribution to AI

Autocomplete was the settled shape of AI coding until March 2024. Tools suggested the next line or answered a question in a chat window, and a human held the loop together. The Devin launch showed something else: an agent given its own sandbox — a shell, an editor, a browser — working unattended for hours, running the tests it had just written, reading the failures, and opening a pull request at the end. The accompanying figure, 13.86 per cent of real GitHub issues resolved without human help against a prior mark near two per cent, mattered less than the demonstrated ability to hold a single task coherent across hundreds of steps. Scepticism followed quickly and much of it was deserved: the freelance-marketplace demo was taken apart in public, the benchmark subset was non-standard, and the company answered with a methodology post rather than a rebuttal. The episode had a useful consequence, in that agent claims afterwards were expected to arrive with reproducible evaluation, and SWE-bench became the scoreboard everyone reported against. The clearer legacy is the scaffold. OpenDevin, later renamed OpenHands, began within days of the announcement and Devika within a fortnight, and the sandbox-with-tools arrangement became the default architecture for coding agents regardless of who built them. Cognition's later argument — that one agent with unbroken context beats a swarm of coordinating ones — remains a live dispute in agent design.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).