Devin
Devin is a commercial coding agent from Cognition, launched in March 2024 and notable for being packaged not as an autocomplete in your editor but as a colleague you assign a ticket to. You describe the task, and it works in its own cloud workspace — a shell, a browser, an editor — reporting back when it has a branch for you to review.
What the product actually provides
The interesting part is less the model than the environment built around it:
- A persistent sandboxed machine per session, so a long task survives across hours and steps.
- A browser it drives itself, to read documentation or reproduce a bug in a running app.
- Integrations with GitHub, Slack and issue trackers — work arrives as a ticket and lands as a pull request.
- A replayable timeline of what it did, which is how a reviewer judges whether the result can be trusted.
Why it matters, and what it taught the field
Devin’s launch demo claimed a far higher resolve rate than the SWE-bench baselines of the time, and the scrutiny that followed became a turning point in how agent claims are reported: benchmark subsets, hidden retries and partial credit all got a lot more attention afterwards. The product itself settled into the pattern the whole category now follows — strong on well-scoped, well-tested changes in a familiar codebase, weak where the task requires judgement about intent. It is best understood as the first serious attempt to sell agentic software engineering as a managed service rather than a library.
Devin is a commercial autonomous coding agent from Cognition, introduced in March 2024, that is assigned whole tasks rather than used as an editor autocomplete: it works in its own sandboxed cloud machine with a shell, browser and editor, and returns a reviewable pull request. Its launch benchmark claims drew heavy scrutiny and changed how SWE-bench results are reported. It marked the shift from code completion to software engineering sold as a managed service.
Most AI coding help works like a fast typist next to you, finishing your sentences. Devin was built to work like a junior colleague instead: you describe a job, it goes away with its own computer, and it comes back with finished work for you to check.
It gets a workspace of its own — somewhere to run commands, a browser to look things up, a place to edit files — and it keeps going over many steps without you steering each one.
When it launched in 2024 the claims about how much it could finish on its own were disputed, and that argument made everyone stricter about measuring these systems. The idea stuck, though: hand over the ticket, review the result.