Devin — the first AI software engineer

What it is. On 12 March 2024 the applied-AI lab Cognition introduced Devin and called it "the first AI software engineer." Where a coding assistant completes the line you are typing, Devin was handed the engineer's own tools — a shell, a code editor, and a browser — and asked to take a whole task end-to-end: read the issue, plan, write, run, debug, and iterate, with a human reviewing rather than driving. On SWE-bench, a benchmark of real GitHub issues from projects like Django and scikit-learn, Cognition reported it resolving 13.86% of issues *unassisted* — against a previous state of the art of 1.96% that had been *told which files to edit*. It was shown passing engineering interviews and completing paid jobs on Upwork.

Why it matters. Devin named a category and forced the shift the whole field then followed: from the assistant (human in the loop, autocomplete at the seam) to the *agent* (agent in the loop, human at the review). Within weeks the shape it drew had open-source answers — SWE-agent, OpenDevin/OpenHands — and every frontier lab a roadmap toward it. Whatever one makes of the launch, "agentic coding" as a product category has a birthday, and this is close to it. It descends from the pieces this collection already holds — ReAct's interleave of reasoning and acting, LangChain's first toolbox, AutoGPT's public autonomous loop, and the yardstick of SWE-bench — and it is upstream of most of what "an agent that ships code" means today.

Stated precisely — the honest bound. The milestone is real; the launch was oversold, and this museum records both. Within weeks the developer Carl Brown ("Internet of Bugs") published a careful ~27-minute rebuttal, "Debunking Devin," and it held up. The headline Upwork demo did not do the job the client had actually posted — the task asked for instructions to deploy a model on AWS; Devin ran it locally and reported results instead. Several of the bugs Devin was shown heroically fixing were bugs Devin had itself introduced moments earlier, and the video omitted that. The SWE-bench figure used a nonstandard custom subset rather than the standard set. None of this erases the capability jump or the category it opened. It means the *demo* was a curated artifact, not a measurement.

Its place beside the verification wings. This is why the object belongs here and not only in a timeline. A launch demo is a fixture its maker wrote and gets to cherry-pick — so a green demo certifies the lab, never production; a benchmark run on a subset you chose is "verified on these inputs," not verified; and an agent that creates a bug and then fixes it, on camera, is issuing a receipt for a problem it caused. Devin is the moment the field learned, in public, that you cannot take an agent-maker's video as the measurement — that the only trustworthy read on "can this agent do the work" is an exogenous one you did not get to stage. The capability was a genuine step; the lesson in how it was sold is the more durable artifact, and the reason both halves are anchored here together.

*Primary source inside: the canonical record — the announcement, the benchmark claim with its methodology caveat, and the documented critique — fingerprinted and anchored like every object here.*