Autopilot: an agent that writes and schedules posts, with no second way to publish

The agent was given no publishing ability of its own. It creates a post, and from there the post goes through the pipeline every hand-written post uses: the same validator, the same approval queue, the same publish scan, retries and calendar. There is no second publish path to secure.

The defaults are the safe ones. Each autopilot is off until someone switches it on, and review is required unless someone turns it off, so generated posts wait as pending approval with a real approval row. Only a manager can configure one, because an autopilot both spends the AI budget and commits content, which is the same authority that approves posts. Every machine-written post carries its own audit action, so it is always distinguishable from a person's.

Failure was designed, not left to chance. A generation error is logged to an append-only run log, the creator is notified, and the slot is skipped, so a topic the model struggles with cannot hammer the API on every scan. An invalid generated post is rolled back whole. Autopilot calls count against the workspace's hourly AI limit like any other use. Topics rotate by run count, and posting slots are spread with jitter inside the team's own working hours and time zone.

A "generate now" button produces one extra post without consuming a slot, which is the sanity check before enabling a schedule.

What this case study does not claim: the feature was verified against a wire-level stub of the Anthropic API, not the live service, so the quality of what the model writes is not yet measured. It also uses one post body for all selected networks and picks images from the media library rather than generating them.