On March 12, 2024, the startup Cognition announced Devin, which it described as “the first AI software engineer.” Unlike a code-completion tool that suggests the next line, Devin was presented as an autonomous agent that plans and reasons over long tasks while operating its own developer environment, including a shell, a code editor, and a web browser, all inside a sandbox. It could report progress in real time and take feedback from a user as it worked.
The announcement claimed Devin could learn unfamiliar technologies, build and deploy applications end to end, debug existing codebases, and even contribute to real production repositories. Its headline result was on SWE-bench, a benchmark of resolving real GitHub issues: Devin resolved 13.86 percent of issues end to end, compared with a prior published state of the art of 1.96 percent, and it did so without being told which files to edit.
Devin became a flashpoint for the whole agent moment. The demo drew enormous attention and helped popularize the idea of an autonomous coding agent, but it also drew scrutiny over how representative the demos were, a debate captured in the separate AI Library story on the topic. For a business reader, the launch marked the point where “an AI that writes a function” gave way to the far more ambitious and contested pitch of “an AI that does the engineer’s job.”