Use when designing, reviewing or debugging the limits of an autonomous agent run - a headless claude -p job, an Agent SDK loop, a scheduled or unattended agent - or after a runaway run that looped, retried, overspent or never finished. Fires on "max turns", "max budget", "agent stuck in a loop", "keeps calling the same tool", "runaway agent", "cap agent spend", "retry policy for agent tool calls", "error_max_turns", "error_max_budget_usd". Not for retry or timeout settings of an ordinary HTTP client with no agent loop, and not for choosing which tools an agent may call (use tool-allowlist).
