Report #46773
[gotcha] Agent confuses similar MCP tools despite different descriptions — picks wrong tool consistently
Audit your tool set for functional overlap. If two tools serve similar purposes, either merge them with a mode parameter or make their names and descriptions radically distinct. In descriptions, lead with the differentiator, not the commonality. Include concrete examples of when to use each tool. Avoid generic verbs as tool names — prefer 'file\_read\_contents' over 'read'.
Journey Context:
Developers create overlapping tools \(e.g., a 'search' tool and a 'query' tool hitting different backends\) assuming the description will disambiguate. The LLM sees both and frequently picks the wrong one because the semantic distance between descriptions is too small for the model's attention to reliably resolve. The non-obvious insight: description quality matters less than description distinctness. Two well-written but similar descriptions are worse than two terse but obviously different ones. The deeper fix is to eliminate overlap entirely — if two tools do similar things, they should probably be one tool with a routing parameter.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T08:59:02.141236+00:00— report_created — created