AI Systems Atlas

G-17Background Agents

/coding-agents/background-agents

Dispatch a task to a remote sandbox, walk away, and come back to a branch and a pull request.

Step-by-step walkthrough8 steps
  1. A developer hands off a task and closes the laptop. Nothing about the run needs them present.
  2. The task lands in a remote sandbox: a fresh clone and the repo’s own setup script, not someone’s machine.
  3. The agent works unattended — explore, edit, and run the checks itself, because nobody else is there to.
  4. The output is not a chat reply. It is a branch and a pull request, and pushing it starts CI.
  5. CI goes red. The failure event, logs included, wakes the agent — nothing polls.
  6. Once the PR is ready, the developer is notified and reviews it when they choose. Hours later is fine.
  7. Review comments are events too: they wake the same agent, on the same branch, to push a fix.
  8. Green, comments resolved. The human spent minutes reviewing instead of hours supervising.
Key insight

The unit of work changes from a conversation to a pull request. The human stops supervising steps and starts reviewing results, and CI failures and review comments, not a watching person, are what wake the agent back up.

Failure mode

Dispatching work the sandbox can’t verify. With nobody watching and no runnable tests, the agent’s only signal is its own confidence, and the PR arrives looking finished.

Further reading
RelatedEnvironment bootstrapCI feedback loopReviewing agent PRsEvent-driven agent

Elsewhere in the atlas

Harnesses →Security →Evals →Context →Tools →Ops →