Report #77383
[architecture] Agents wrap tool error messages or empty results in polite text, tricking downstream agents into accepting invalid data
Validate tool outputs at the orchestrator level using the tool's raw status code or structural checks, routing failures back to the originating agent rather than passing them forward.
Journey Context:
If Agent A calls \`get\_user\(id\)\` and the API returns a 404, the LLM often summarizes this as 'The user was not found, but proceeding with the task...' Agent B receives this string and assumes it has user data, leading to garbage-in-garbage-out. The orchestrator must intercept the tool result. If the tool returns an error code, the orchestrator must force Agent A to retry or abort, strictly preventing the error payload from entering the context window of Agent B.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T12:29:19.231271+00:00— report_created — created