← Integration Engineering
Integration Engineering

Too Many Tools: Why a Bigger Toolbox Makes an AI Agent Worse

22 Oct 2026·3 min read

Connect an AI agent to a CRM, a ticketing tool, a calendar and an HR system. Each brings forty or fifty actions. Now the agent has a few hundred tools, and it starts making strange choices.

What goes wrong

  • Tokens: every tool has a name and description the model reads on every request. OpenAI’s guide says these count against the context limit and are billed as input tokens.
  • Wrong picks: more tools mean more wrong choices. The same guide suggests aiming for fewer than 20 at the start, as a soft suggestion. Tools from different apps, like create_ticket and create_issue, look alike.
  • Limits: some model APIs cap how many tools one request can carry. Check your model’s current docs.

The fix: search first, then run

Give the agent two tools instead of hundreds. One searches the catalog of actions in plain words. The other runs the action it found. Only what is relevant enters the context.

[
  { "name": "search_actions", "description": "Find actions by describing the task in plain words", "input": { "query": "string" } },
  { "name": "run_action", "description": "Run one action returned by search_actions", "input": { "action_id": "string", "params": "object" } }
]
Example shape, not a real API. The agent sees these two tools no matter how big the catalog is.

If your agent does one focused job with ten tools, show it all ten. Search pays off when the agent spans many apps.

Do this in 60 secondsCount the tools your agent gets at the start of a request. If it is well over twenty, remove the ones it never uses and rename pairs with confusing names.
Today’s AI angleEvery app you connect adds tools, so your tool list grows with your customer’s stack. Plan for search early, and log which tool the agent picked so you can see the wrong picks.
Too Many Tools: Why a Bigger Toolbox Makes an AI Agent Worse · Ali Huzaifa