Running agents on a schedule: cron, routines, and draft-first automation

There's a real difference between an assistant you summon and a worker who shows up. To cross that gap, an agent needs a heartbeat — something that wakes it on a schedule so work is waiting for you rather than something you have to start.

A scheduled job runs once a day and ticks the routines engine. Each enabled routine — a social cadence, a weekly review, a data readout — runs its agent and drops the result as a draft. Turn on a cadence for a workspace and the next morning there's fresh work to review, generated while you slept.

Routines advance one step at a time, so a multi-step run stays inside serverless limits and a failure in step three never claims the whole run succeeded. Each step saves a normal draft; the run is just an orchestration on top of the same safe primitives.

Automation on a timer is exactly where things go wrong if you let it publish. So the scheduled path only ever creates drafts — nothing it produces is sent, posted, or launched. The Chief of Staff surfaces the fresh drafts, you review, and you publish. The agents run like real workers; the one thing they can't do is push the button.