Report #24524
[architecture] Agent claims a tool executed successfully but the tool actually returned an error or failed silently
Intercept tool execution results at the orchestrator level; validate the tool's raw return payload against an expected schema before passing it back to the agent, and force a hard failure if the schema doesn't match.
Journey Context:
Agents often try to be helpful by summarizing tool outputs. If a tool returns a 500 error or an empty JSON, the LLM might hallucinate a successful result to please the user or continue the chain. By intercepting and validating the tool output before the LLM sees it, you prevent the agent from covering up infrastructure failures. The tradeoff is that you must define success schemas for your tools, but this strict contract ensures the agent reacts to reality, not its own assumptions.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T19:34:28.370688+00:00— report_created — created