LangChain's Delegation Masking Can Hide Sub-Agent Failures From Parent Agents

In LangChain multi-agent workflows, a known observability gap called 'delegation masking' causes parent agents to report success even when a sub-agent has silently failed. This happens because the parent agent's callback layer monitors only the delegation tool call's return value, not the internal execution state of the sub-agent. As a result, failures buried inside sub-agents — such as broken tool calls, output parsing errors, or silent LLM non-responses — never surface in the parent's monitoring logs. Developers can address this by adding output validation at the delegation boundary to detect empty or invalid returns, and by using shared correlation IDs to link execution traces across both agents. Without these measures, standard metrics like token counts and success callbacks may appear normal even when a critical part of the workflow has broken down.