Report #11345
[gotcha] LLM tool selection accuracy degrades severely with 50\+ registered tools
Group tools into distinct namespaces or semantic categories, and only expose a subset of tools per turn, or use a router agent to select the tool subset.
Journey Context:
It is tempting to expose a massive monolithic MCP server with every possible tool. However, LLM attention mechanisms dilute over long tool lists, causing them to confuse tools with overlapping functionality \(e.g., search\_code vs search\_file\). By using a hierarchical approach or a tool-routing step, you drastically improve selection accuracy at the cost of an extra LLM call, which is a worthy tradeoff.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T13:09:38.745656+00:00— report_created — created