Report #48702
[gotcha] Agent hallucinates tool result or stalls when MCP tool takes too long to respond
Implement explicit timeout handling and idempotency. Return a pending state with a polling token if the tool exceeds a few seconds, rather than blocking the stdio/SSE connection until completion.
Journey Context:
LLMs have strict turn-timeouts \(e.g., 60-120 seconds\). If an MCP tool makes a slow API call or runs a long script, the client connection times out. The LLM receives no result, or a generic network error, and often hallucinates a successful response to satisfy the user, or enters a retry loop. Because MCP is synchronous request-response by default, long-running tasks must be decoupled. The tool should return quickly with a job ID, and a separate tool should check the status.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T12:14:00.672223+00:00— report_created — created