AGT-09 · Agent failure & recovery¶
When an agent loops, hits a tool outage or stops halfway through a multi-step task, teams find out from the cloud bill or an angry user. Each project invents its own retries and nobody knows how to resume, undo or replay a broken run.
- Who reads it: Agent engineers, GenAI engineers and on-call DevOps engineers running LangGraph agents in production
- Outcome: Loops stopped in minutes, failed runs parked, resumed or undone, never lost
- Pack owner: Sivaraj (catalog owner role: Agent Lead)
Master catalog entry¶
- Catalog name: Agent Failure & Recovery Runbook (P2)
- Why it matters: Graceful failure preserves trust; silent loops destroy it
- What it removes: Loop detection, timeout and compensating-action handling standardised
- Stack: LangGraph, OpenTelemetry, dead-letter queues
- Maturity target: Level 2 — covers infinite loops, tool outage and partial completion
Playbooks¶
| ID | Story | Title | Kind | Complexity | Urgency | Status | Owner | Planned publish |
|---|---|---|---|---|---|---|---|---|
| PB-AGT-09-01 | AGT-09-S1 | Stop and triage an agent stuck in a loop or burning tokens | Runbook | Medium | 5 | draft | Sivaraj | 2026-11-25 |
| PB-AGT-09-02 | AGT-09-S2 | Handle a tool or model outage during agent runs | Runbook | Medium | 5 | draft | Sandeep Velu | 2026-11-30 |
| PB-AGT-09-03 | AGT-09-S3 | Build compensating actions and checkpoint resume into a LangGraph agent | How-to | High | 4 | draft | Sivaraj | 2026-12-08 |
| PB-AGT-09-04 | AGT-09-S4 | Drain and replay the agent dead-letter queue | Runbook | Medium | 4 | draft | Sandeep Velu | 2026-12-14 |