Error Recovery and 503 Failures

• Classify errors before retrying.
• Design retries with backoff, jitter, and limits.
• Escalate to human review without losing state.

Classify Before Acting

• Transient: 429, 503, timeout, broken connection. Retry with backoff.
• LLM-recoverable: invalid JSON format, missing field. Retry with explicit correction.
• User-recoverable: missing data, permission, or approval. Ask or pause.
• Permanent: invalid credential, resource does not exist, policy blocked. Stop and report.

type AgentError =
  | { kind: "transient"; code: 429 | 503 | "timeout"; retryAfterMs?: number }
  | { kind: "model_format"; message: string }
  | { kind: "needs_human"; reason: string }
  | { kind: "fatal"; reason: string };

const retryPolicy = {
  maxAttempts: 4,
  initialDelayMs: 1000,
  backoff: 2,
  jitter: true,
};

Minimum State to Avoid a Spiral

{
  "task_id": "research-2026-07-05-001",
  "step": "fetch_sources",
  "attempts": 2,
  "last_error": "503 from search API",
  "next_retry_at": "2026-07-05T10:31:00Z",
  "fallback": "use cached sources",
  "human_review": false
}

Recommended Recovery Pattern

• Execute the tool with a timeout.
• If it fails, classify the error in code, not in free text.
• If transient, wait with backoff and log the attempt.
• If attempts are exhausted, switch to fallback or escalate to a human.
• Summarize the state, not the full history, to continue.

Official Sources

• LangGraph persistence
• LangGraph: Thinking in LangGraph
• LangChain/LangGraph human-in-the-loop