Too Many MCP Tools: I Gave My Agent 61 and It Got Dumber

Описание и применение

Слишком много инструментов MCP ухудшают точность выбора инструментов агентом. Ограничив активные инструменты до 10‑15, используя под‑агентов и обёртки, можно восстановить точность и снизить токен‑трассировку.

Задача

An AI agent with too many Multi‑Channel Platform (MCP) tools mis‑selects the correct tool, leading to repeated wrong calls, context pollution, and reduced task accuracy.

Результат

Limiting active tools to 10‑15 per agent and structuring them by domain or intent restored first‑call accuracy to ~90% and eliminated context pollution caused by repeated wrong tool calls.

Рабочий процесс

• Identify and log all connected MCP tools.
• Delete unused servers and tools.
• Allowlist only essential tools per server.
• Create sub‑agents for distinct domains (code, docs, comms).
• Wrap multiple related tools into a single intent‑based tool.
• Monitor first‑call‑correct rate and tool usage logs.
• Iterate to keep tool count within 10‑15 per agent.
• Use dynamic tool loading if supported to reduce prompt tax.
• Adjust tool descriptions to match prompt vocabulary.
• Re‑evaluate accuracy after changes.

Доказательства

First‑call‑correct rate went from ~92% at 12 tools to ~61% at 61 tools on the same task set.

Источник

The fix is a tool budget: 10‑15 tools per agent, scoped sub‑agents, and wrappers that collapse a whole server into 2‑3 verbs.

Источник