Report #49794
[architecture] Infinite loops where agents pass tasks back and forth or evaluator keeps rejecting
Enforce a strict max\_iterations counter on loops \(generator-evaluator or agent handoffs\) and a circuit breaker. If the limit is hit, escalate to a human or fallback logic rather than retrying infinitely.
Journey Context:
LLMs are stateless and can get stuck in repetitive failure modes. If an evaluator says 'missing X' and the generator keeps failing to provide X, they loop forever. A hard limit with an escape hatch prevents runaway API costs. Tradeoff: Might abort a task that would have succeeded on iteration 6, but 5 is a reasonable heuristic for most LLM loops.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T14:03:39.793497+00:00— report_created — created