Report #5108
[gotcha] Agent assuming tool execution succeeded when it silently failed or returned an error
Implement robust telemetry and logging for all tool invocations, including latency, status codes, and error messages. Ensure the LLM is explicitly prompted to check for error indicators in tool returns and halt if a critical tool fails.
Journey Context:
MCP servers can timeout or return error objects. If the agent host doesn't parse the error or log it, the LLM might receive a generic error string and hallucinate that the operation succeeded, proceeding with flawed logic. Without telemetry, developers are blind to silent failures causing cascading errors in agent workflows.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T20:40:37.442979+00:00— report_created — created