What it is
Devin is an agent from Cognition, presented as "the first fully autonomous AI software engineer": you give it a task, and it plans, writes code, debugs, deploys in its own environment and hands back the result. That framing set the bar — and the vocabulary — the whole market then used to talk about autonomous coding agents.
Where it came from
- March 2024 — Cognition announces Devin with a demo of it "solving real jobs on Upwork". Huge resonance, and plenty of argument.
- The demo triggered a wave: "if an agent can do that, where are we in a year?"
- 2025 — the fire-and-forget format was picked up by Codex Cloud, Cursor background agents, Copilot coding agent and Google Jules.
Why it took off
- It shifted the frame from "assistant" to "doer": you delegate the whole task to the agent, like to an employee.
- It gave the industry a shared narrative and metrics (share of issues resolved, PRs accepted).
- It provoked a wave of sceptical teardowns — which also kept the topic in public view.
How to use it today
- Use an "engineer agent" where the task is isolated and verifiable (a bug with a repro, a routine feature).
- Treat the output like a PR from a junior: review and tests are mandatory.
- Compare tools by benchmarks and by your own repo, not by demos.
What to watch out for
- The original "first AI engineer" demos were criticised for embellishment — separate marketing from reproducible results.
- Full autonomy on real legacy code is still fragile; the longer the chain, the higher the chance of a dead end.
- Delegating without guardrails (permissions, sandbox) is dangerous — see agent security.