Agent runbooks

The Operator

The Operator runs every 2 hours. It checks every MCP connection and every agent schedule, routes unassigned work, and escalates anything it cannot resolve. It takes no direct external actions.

What the Operator monitors

  • -MCP connection health for all registered sources.
  • -Agent schedule health: whether each scheduled agent ran as expected.
  • -The event log for errors and unrouted work items.
  • -All agent_installs health status.

Health check workflow

  1. 1

    Check each MCP source

    The Operator calls a lightweight probe on each registered MCP connection. If the source responds normally, it logs a health event and moves on.

  2. 2

    Attempt reconnect on degraded sources

    If a source is degraded or offline, the Operator runs the reconnect runbook for that source. For many transient failures, this resolves the issue automatically.

  3. 3

    Escalate what it cannot resolve

    If the reconnect fails, an urgent escalation card is created with the exact steps to fix the issue manually. The card includes the error message, the expected state, and the commands to run.

  4. 4

    Route unassigned work

    The Operator checks the event log for work items that no installed agent is configured to handle. These become escalation cards.

Schedule

Every 2 hours.

The every-2-hours cadence means a connection issue is detected within two hours of it occurring. In most cases, the Operator resolves transient failures automatically. Persistent failures appear in your inbox as escalation cards within two hours.

No data sources required

The Operator reads only internal FlostackOS state: MCP connection health, agent schedules, and the event log. No external MCP connections are required for the Operator to run.

Troubleshooting guides

The Operator uses the same reconnect steps you can run manually. If you receive an escalation card from the Operator, you can also follow these guides: