Devin

Devin’s design question is what happens when you remove the human from the inner loop entirely. The agent gets a workspace with a shell, an editor, and a browser, and works through a task on its own timeline. Output is a pull request, not a diff you approve line by line.

The economics of asynchronous work

Supervised agents cost engineer attention proportional to run time. You watch, you approve, you correct — and the saving is bounded by the fact that you are still there.

An asynchronous agent costs a fixed review at the end regardless of how long the work took. Assign five tasks in the morning, review five pull requests in the afternoon. That is a genuinely different operating model, and for a backlog of small well-specified work it is the first thing in this category that changes the shape of a team’s week rather than the speed of an individual’s hour.

The workspace matters to making this credible. A browser means the agent can read documentation and reproduce a bug in a running application. A persistent shell means it can install dependencies, run migrations, and iterate against a real environment rather than reasoning about one.

Where the model breaks

Errors compound silently. An early misreading of requirements propagates through every subsequent commit, and you only see it at review — by which point there is a large, internally consistent, entirely wrong pull request to evaluate. Reviewing that costs more than reviewing a correct one, and much more than catching the misreading in the first minute.

This inverts the usual advice about specification. With a supervised agent, a vague task is recoverable because you correct it as it goes. With Devin, the specification is the only steering input you get, and a vague one reliably wastes an entire run.

The practical discipline is to write acceptance criteria before assigning, not requirements. “The endpoint returns 429 with a Retry-After header when the per-key limit is exceeded, and there is a test proving it” steers a run. “Add rate limiting” does not.

Cost and predictability

Subscription with usage-based consumption on top, no free tier, and enterprise plans available. The consumption model means a run that flounders costs real money and produces nothing, which makes task selection a financial decision rather than only a technical one.

Success rate varies enormously with task type, and the honest way to evaluate is to run a batch of ten representative tickets and count. Teams that pilot with their easiest work and extrapolate get a badly wrong number.

Where it fits

Backlog work with clear acceptance criteria — dependency bumps, adding tests, small well-specified features, migrating patterns already established elsewhere in the codebase. Work that a competent junior could do from a well-written ticket, in other words, and where the ticket was actually well written.

Architectural work and anything requiring taste is still a poor fit, as is any change where “correct” depends on context that lives in someone’s head rather than in the repository.

Teams that want most of the leverage with a human checkpoint should use Claude Code or Cursor Agent instead, which is the right default for the large majority of engineering organisations today.

Compare it against supervised alternatives in our best AI agents roundup.