The rule
Accounts die from silence, not from bad posts. The engine's job is a guaranteed daily pulse per client: content going out, engagement going back, failures surfacing loudly instead of rotting quietly in a queue.
When to run it
Continuously. This is the always-on layer under everything else — scheduled runs several times a day, plus a morning health sweep.
The play
1. Cron the publish queue at minute resolution; every failed publish gets logged with the platform error, not swallowed. 2. Morning sweep: one automated report — queue depth, failures overnight, tokens near expiry, accounts with nothing scheduled in the next 72 hours. 3. Engagement rounds a few times daily: reply to comments, like relevant activity, follow up on DMs — bounded per round so it reads human, never spammy. 4. Auto-retry transient failures once; anything that fails twice escalates to a human with the exact error and a one-click retry. 5. Weekly: a plain-English client note (three sentences and the numbers). Written by the agent, sent on schedule, readable by a business owner in transit.
Where it breaks
OAuth token expiry is the silent killer — surface expiry dates proactively, because platforms won't. And never let the agent use a client's connected account to promote your own products; the engine serves the client, full stop.