Pattern Automation
← Learn glossary

Learn

What is agent failure recovery?

Agent failure recovery is how roles resume after tool errors, model timeouts, bad outputs, or human rejection — with checkpoints, retries, fallbacks, and escalation instead of silent failure or infinite loops.

In depth

Agents fail. Connectors rate-limit, models hallucinate, humans reject drafts. Recovery design answers: what state is saved, what retries, when to escalate.

Recovery stack:

1. Idempotent steps — safe to retry without double send (use draft + Ask) 2. Checkpoint — task status and partial artifacts persisted 3. Bounded retries — exponential backoff on transient errors 4. Fallback path — route to human queue with full context 5. Post-incident skill patch — failing eval case added to regression set

Without recovery, operators babysit chat sessions. With recovery, heartbeats pick up from `blocked` state next run.

Neuro OS run receipts show failure class — timeout, tool 403, eval fail, human reject — so fixes target root cause.

Examples

  • CRM write fails — role saves draft; retries after token refresh
  • Human rejects email — task returns to copy role with rejection note
  • Model timeout — heartbeat retries once; then escalates
  • Poisoned RAG chunk — input guardrail blocks; incident logged

Related terms

FAQ

How many retries?

1–3 for transient errors; never retry external send without idempotency.

Should agents self-heal silently?

Internal retries yes; external actions need Ask after material change.

Run governed agent roles on a company OS — not only definitions in a glossary.