Skip to main content

The Tool Junk Drawer

Exposing the agent to a large number of tools — often thin wrappers around every available API endpoint — with vague descriptions, overlapping functionality, and no strategic curation. Treating tool quantity as a proxy for agent capability.

Why It Happens​

  • More tools feels like more capability
  • API-wrapping is easy to automate
  • Teams want to be "complete" and worry about missing a use case
  • MCP servers make it trivially easy to expose dozens of tools
  • Removing a tool feels like removing functionality

What Goes Wrong​

  • Selection accuracy collapses — research shows accuracy drops to as low as 13% with large toolsets
  • Decision paralysis — overlapping tools create confusion about which to invoke
  • Context bloat — every tool description consumes tokens in every request
  • Poor error recovery — tools with cryptic error codes leave agents unable to self-correct
  • Wasted context — tools returning excessive data waste the context window

What to Do Instead​

  • Curate ruthlessly — Anthropic says "Don't wrap every API endpoint as a tool. Focus on high-impact workflows"
  • Consolidate — merge multi-step operations into single tool calls
  • Clear descriptions — treat tool descriptions as prompts, not documentation
  • Human-readable output — return readable context, not raw IDs or technical dumps
  • Actionable errors — error messages should tell the agent what to try next
  • Tool RAG — for large toolsets, dynamically retrieve relevant tools per query instead of loading all

Signs You Have This​

  • Agent has access to 15+ tools
  • Multiple tools do roughly the same thing with slightly different names
  • Tool descriptions are copy-pasted from API docs
  • The agent frequently calls the wrong tool or calls tools that return errors
  • You've never removed a tool, only added them