All posts
6 min read

Cron for AI agents: scheduling work when nobody's watching

The whole point of an always-on agent is work that happens without you. The mechanism is old and boring on purpose: cron. A schedule line, a prompt, a target agent — and one genuinely hard question the cron part doesn't answer: what should an agent be allowed to do when there is nobody around to say yes?

A schedule is a cron line with a target agent

On Choreo, a schedule is a cron expression (evaluated in UTC — presets from "every weekday at 8am" down to every minute for testing), a prompt, and a choice of which agent on the project runs it. Each schedule can also override the model for just that job — a cheap model for the hourly heartbeat, a stronger one with more reasoning effort for the weekly analysis. Where the result goes is a four-way choice:

  • Announce — the reply posts to a messaging channel (WhatsApp, Telegram, Slack…), like the agent messaged you first.
  • Web — the result is stored in the dashboard, no channel needed. Fire-and-review.
  • Webhook — the trigger metadata POSTs to your URL, for wiring agents into systems you already run.
  • None — it runs internally. Useful as a heartbeat or for work whose output is the side effect.

Nobody's there to click "approve"

Interactively, capabilities are allow / ask / deny — and "ask" means a human answers a prompt. A scheduled run has no human. Anything set to allow runs on a schedule exactly as it does interactively — web search is allow by default, which is how a scheduled agent reads the web at all. The rule is about the middle setting: anything set to "ask" is denied on a scheduled run, unless you pre-approved that specific capability for that specific schedule. Pre-approval is revocable, covers only mild capabilities (writing files, running code, running shell commands) — and, quoting the product copy directly: "email sending, arbitrary APIs, file deletion, credentials and computer control are never grantable unattended." The denials are the feature. A scheduled agent that drafted an email will be holding it for you in the morning, not explaining at 3am why it sent it.

Schedules that drive a browser

A scheduled run can open a real browser — an isolated virtual desktop with Firefox that the agent operates by looking at the screen. Computer mode is switched on per schedule, and with session persistence on, the same schedule reuses the same browser profile fire after fire, so logins and cookies survive between runs. This is the shape of "check this dashboard every morning at nine": a page that needed a login on Monday just works on Tuesday. Everything the agent does in there is still governed by the project's capability policy — and if that policy says "ask" for computer control, the unattended run is refused rather than silently approved. The thing that deliberately does not exist is a standing grant that turns an unattended "ask" into a yes.

Failure is handled, not hidden

Schedules fail in two ways, and they're treated differently. A transient failure — a rate limit, a network blip — just waits for the next tick. A permanent one (a misconfigured prompt, a revoked connection) is counted, and after repeated consecutive permanent failures the schedule disables itself and tells you, instead of failing identically every morning forever. And because every model call, tool call and connector request is recorded append-only with its cost, an unattended run is the most auditable kind there is — the full trail is described on the trust page. If you have one recurring job you keep doing by hand, that's the schedule to create first — free plan, no card, and the strict unattended rules above are the default, not an upgrade.

See it for yourself — free, $2 of credits once you confirm your email, no card.

Start free