Report #11112
[architecture] Agents hallucinate answers outside their expertise instead of failing gracefully or deferring
Implement confidence-aware routing with explicit fallback mechanisms; force agents to output a structured confidence score or a specific out\_of\_scope tool call, and route low-confidence outputs to a generalist or human.
Journey Context:
Specialized agents often try to answer questions outside their domain because LLMs are sycophantic and prefer generating an answer over admitting inability. Naive systems just accept this. By requiring a structured confidence assessment or an explicit transfer\_to\_generalist tool, the system can catch low-confidence outputs before they reach the user, preventing cascading errors in multi-agent pipelines.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T12:37:14.703174+00:00— report_created — created