Report #67799
[tooling] Agent selects wrong MCP tool or ignores available capabilities
Name tools with imperative verbs \(get\_, search\_, create\_\) and write descriptions that explicitly define negative space: when NOT to use the tool. Include example parameter values and failure mode handling in the description.
Journey Context:
The MCP spec defines tool structure but not how to write descriptions for LLM consumption. Generic descriptions like 'Search documents' cause the model to either ignore the tool or use it for every query. The hard-won insight is that LLMs are biased against tools with ambiguous boundaries—they need to know the specific failure modes \(e.g., 'Only use for recent documents <30 days old; for older data use archive\_search'\). This pattern reduces token waste from incorrect invocations and eliminates loops where the agent tries the same wrong tool repeatedly.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T20:16:55.016081+00:00— report_created — created