Operating Fleets

Day-2 operations for fleet engineering.

Weekly

• Review FLEET-STATE.md — retire stale agents
• Check budget burn vs caps
• Inbox queue depth for risky action classes
• One accountability test spot-check

Monthly

• Re-run fleet-audit on each team workspace
• Update patterns/registry.yaml if new patterns adopted
• Review clone/fork sprawl
• Kill switch drill (pause one non-critical agent)

Logging

Minimum fleet log fields:

timestamp, agent_id, version, principal, action, task_ref, trace_id, outcome

Store where compliance requires — not only in chat UI.

When to kill

Kill or pause when:

• Budget hard cap hit
• Accountability test fails on production action
• Inbox SLA breached for destructive class
• Agent version drifts from registered manifest
• Incident commander requests FLEET_PAUSE_ALL

Document kill in FLEET-STATE.md with reason and owner.

Cost operations

Pair with loop-engineering operating loops:

Loop ops	Fleet ops
Per-loop token estimate	Per-agent + per-team caps
loop-run-log	Fleet-wide spend dashboard
Kill one loop	Kill agent subset or fleet

Promotion workflow

F0 experiment → manifest draft → F1 catalog → inbox trial → F2 shared → F3 enterprise

Never promote on demo success alone — require accountability test pass for the action class.