Deterministic checks, not careful prompting. Loop detection, credential scanning, isolated branches for automated work, and a supervisor that keeps the daemon alive for weeks unattended.
A monitor watches tool calls over a rolling window for the signatures of being stuck: the same call returning the same result, a failure retried unchanged, a check failing the same way, two actions oscillating. It intervenes rather than letting it run.
What that gets you. You are not billed for an agent spinning on the same mistake, and it cannot quietly loop all night.
Every dispatched goal runs in its own git worktree on its own branch. Finished work is credential-scanned and pushed to that branch, never to main. Failed attempts push their partial work too.
What that gets you. Autonomous changes cannot touch your main branch, and no attempt's work is ever silently lost.
A circuit breaker forces a clean restart rather than letting the daemon limp along damaged, and an external watchdog restarts it if it stops responding while still technically running.
What that gets you. Scheduled work still happens in week three, which is the only thing that makes background automation worth setting up.
A dedicated agent on its own schedule sweeps your repositories to draft commit messages, PR descriptions and changelogs, and to spot stale merged branches and orphaned worktrees.
What that gets you. The tidying you keep meaning to do, done on a schedule you did not have to think about.
Everything on this page describes behaviour that ships today. The facts come from the runtime’s own capability brief, the same one the agent reads about itself on every turn, so nothing here is a roadmap item wearing a present tense.
How does an AI agent remember across sessions?
Can an AI agent run without sending your data to the cloud?
Can an AI agent work while you sleep?
What can an AI agent actually do on your computer?
How do you stay in control of an autonomous agent?
Which AI models can you use, and who pays for them?