An agent fleet that pages itself

Eighty agents across eighty containers, one per customer. Each runs its job. Each writes a heartbeat row to a coordinator container's SQLite once a minute.

A cron in the coordinator runs every two minutes: count agents missing recent heartbeats, group by error type. If more than three are silent, it POSTs to hoody-notifications: urgency: critical, summary: "7 agents silent, errors clustered around: rate-limit-anthropic". If exactly one is silent, urgency: normal with the customer ID.

The fleet polices itself. The notification only fires when something's wrong, with the diagnosis already inside it. You haven't built a dashboard. You haven't subscribed to anything.

Per-customer agents on hyperscalers run a meter the whole time they exist. Here they run on the box you already rent. The notification layer rides on the same box.

The fleet noticed itself was broken before you did.