Update
Agent Loops: an agent that decides when to wake up next, inside limits you set
TL;DR
Agent Loops are agents that wake themselves up instead of running on a fixed cron: a loop has an objective, stop conditions and limits, and after each run it decides when to wake next or that it is done. Loops live on the Schedules page's Loops tab.
What a loop is
- An objective and stop conditions — what the loop is working towards and when it should end.
- Limits: the number of iterations, a total runtime of up to 180 days, a credit budget, the shortest and longest wait between wakes, and an end date.
- After each run it decides when to wake next, within the wait window, or that it is done. Limits, credits, repeated failures and a person's stop end it deterministically — the model does not get a vote on those.
- Each run is an ordinary task in Task history. A run waiting for a person holds the loop until someone answers.
Control and approvals
A loop can be paused, resumed, cancelled or woken now. Every change a loop's run would make waits for approval by default; a loop's own auto-approve is opt-in. Each project's overseer is a loop of this kind.
Who sets up a loop (updated 20 October)
- A person — an admin or editor — on the Schedules page's Loops tab.
- A project — its overseer is a loop, created when the project is activated.
- The agent itself, from a task — ask it to keep working on something ("keep optimizing until CPA is under 25") and it can propose a loop. Creating one shows an approval card by default (unless that task has auto-approve on); the loop it creates always has auto-approve off, and everyone behind the task must be an admin or editor.
- What the agent cannot do: a loop's own run never starts another loop, and a project's task adds a step to its plan instead. The agent can also list loops and pause, resume or stop one — each change through the same approval.
Frequently asked questions
- What is the difference between a schedule and an agent loop?
- A schedule runs on a fixed cron. A loop has an objective and limits, and after each run decides when to wake next — within the wait window you set — or that it is done.
- Can the agent set up a loop by itself?
- Yes, from a task: when you ask it to keep working on an outcome, it can create a loop. By default that waits for your approval card (unless the task auto-approves), the loop's own auto-approve is always off, and a loop's run can never start another loop.
- Can an agent loop run forever?
- No. Every loop has limits — iterations, a total runtime of up to 180 days, a credit budget and an end date — and those, repeated failures or a person's stop end it.