Running Buzz Agents Reliably and Continuously

Continuous operation depends on both Buzz's relay/harness lifecycle and the host's provider, scheduler, credentials, resources, and persistence controls.

Reliable operation compares desired and observed state, wakes only due work, records heartbeat and job receipts, reconciles bounded fields, and proves repairs after restart.

1. 1Record desired identity, harness, model, prompt version, tools, trigger, and parallelism.
2. 2Observe those fields from the live runtime.
3. 3Run a heartbeat that is a no-op when nothing is due.
4. 4Persist failures with an immutable fingerprint and repeat count.
5. 5Reconcile one bounded change and perform delayed post-restart read-back.

• No-op heartbeat does not restart healthy agents.
• Scheduled work produces a run event and artifact.
• A deliberate drift repair survives restart and delayed recheck.
• Repeated blockers become a permanent capability, governance, or dependency fix.

• A separate launch guard restores stale configuration.
• Heartbeat reports green while child commands fail.
• Auto-wake starts expensive agents with no due job.
• Jobs disappear across restart because state exists only in memory.

Add cost controls after lifecycle correctness is proved; cheap failure is still failure.

• Buzz Agent costs
• Buzz Agent workflows
• What Is Buzz?
• Buzz Agent troubleshooting