Report #74483
[architecture] Upstream agent misinterprets a tool error or empty result as a successful payload, passing malformed data to the downstream agent
Implement a deterministic validation gate on tool outputs before the LLM processes the result, and inject a hardcoded error message into the context if validation fails.
Journey Context:
LLMs often try to recover gracefully from bad tool outputs, masking the failure from downstream agents. For example, an API returning an HTML error page might be summarized as no data found. By intercepting the tool output programmatically and replacing it with an unambiguous system message, you force the agent to retry or abort rather than passing garbage down the chain.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T07:37:05.715428+00:00— report_created — created