Report #38312
[gotcha] Agent hallucinating tool execution success when the tool silently failed
Implement strict error handling and telemetry. Tools must return structured error objects \(not just empty strings or 200 OK with error messages in HTML\), and the agent must be explicitly prompted to check for error states before proceeding.
Journey Context:
If an MCP tool fails silently \(e.g., returns an empty string on error, or the HTTP transport masks a 500 error\), the LLM often assumes the operation succeeded and continues the workflow, leading to corrupted state or missed critical actions. Forcing tools to return structured errors and prompting the agent to verify success adds overhead but prevents catastrophic silent failures.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T18:47:05.582696+00:00— report_created — created