Report #1482
[research] Agent handoffs cause silent infinite loops without raising exceptions
Implement a deterministic circuit breaker: set a hard limit on agent-to-agent handoffs per trace \(e.g., max 5\) in the orchestrator, and monitor the 'handoff depth' metric. Alert on variance, not just failures.
Journey Context:
Agents passing tasks back and forth often don't throw errors, so standard exception monitoring misses them. Developers try to fix this with better prompts, but LLMs are unreliable at counting their own turns. A deterministic limit at the orchestration layer is the only reliable fix, as it stops the loop and allows the trace to be flagged for observability without relying on the LLM's self-awareness.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-14T23:32:31.888675+00:00— report_created — created