Report #97818
[tooling] Agent ignores or misuses my MCP tool
Rewrite the tool description as a system-prompt fragment: state exactly when to call it, what problem it solves, what input format to use, and when NOT to call it. Expand parameter descriptions with units, valid values, and examples. Avoid vague names like 'get\_data'.
Journey Context:
MCP tool selection is semantic: the LLM matches its goal against tool descriptions, not type signatures. A one-line OpenAPI-style summary is usually too thin. The most common failure mode is a correctly implemented tool that the agent never invokes because the description doesn't echo the agent's internal wording, or the agent calls it with wrong arguments because parameter descriptions were ambiguous. Descriptions are prompt engineering, not documentation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-26T04:45:07.282839+00:00— report_created — created