Until now every task was a one-shot: submit, get a PR, repeat. Two new capabilities turn that into a standing fleet you can point at an objective and govern.
Recurring workflows (/loop)
Type /loop in the composer and hand the agent a standing objective instead of a single task:
/loop keep the test suite green and fix new type errors
/loop 6h triage new bug issues and open a fix PR- Self-paced loops work toward the goal and stop on their own once it's provably met.
- Scheduled loops re-run every interval (
30m,6h,1d).
Each iteration runs as a full autonomous goal — plan, code, test, PR — and you stay the merge gate. The loop shows up as a live card (iterations, spend, next run, every PR) and a cross-session Workflows view; pause, resume, or stop any time.
Every loop is bounded — a max iteration count and a total-budget cap on top of the per-task cost cap — so an unattended fleet can never run away. Loops can also be stood up automatically from a recurring trigger, pointed at a project's objective, so ongoing work keeps moving without anyone kicking it off.
Bring your own guardrails (hooks)
On top of the guardrails that already hold — branch protection, cost cap, egress denylist, scrubbed logs — you can now add your own policy with hooks, in Settings → Hooks. They're enforced server-side at the points an agent can't route around:
- block a prompt that hits your egress denylist or leaks a secret,
- gate which MCP servers and models an agent may call,
- notify your own webhook the moment a goal finishes.
Security hooks fail closed, any secret you give a hook is encrypted at rest, and nothing runs until you add one.
Why
The whole point is to hand off outcomes and walk away. Loops let you hand off an ongoing outcome, and hooks let you set the rules the fleet operates under — without giving up the per-task caps and human merge gate that keep it safe.