An Agent With Every Tool Is Worse at Using Any

The cost problem

Every tool you connect comes with a description and a full parameter schema, and all of it is sent to the model on every single request. Not once at setup. Every message. The model keeps no memory of your tool list between turns, so the whole catalogue goes along for the ride each time.

Forty tools with a paragraph each means you ship forty paragraphs before the user has finished typing hello.

Cloudflare ran this on their own API. As tool definitions it comes to 1.17 million tokens, more than any current model can hold. The tool list alone does not fit in the context window, and no work has happened yet.

That is the cost problem, and it is the one everybody notices.

The bigger problem is accuracy

A model choosing between twelve relevant tools chooses well. The same model scanning a thousand does not. Anthropic measured it. Tool selection accuracy went from 49 percent to 74 percent on one benchmark, and 79.5 to 88.1 on another, purely by not loading everything upfront. Same model, same tools available. The only change was how many it had to read at once.

So the agent that has everything is worse at using anything.

Why more agents isn't the fix

The obvious fix is to split the work across several agents with smaller tool sets. It works, but it costs more than it looks. A second agent does not know the conversation, so you brief it, and what comes back is a summary. You get a report of what it meant to do rather than what it did. Sub agents are a good way to isolate a large messy result. They are a poor way to isolate a tool list.

Load only what's needed

What actually works is simpler. Do not load a tool until it is needed.

The agent sees only the names. When a request comes in it searches its own catalogue, pulls the two or three definitions that matter, and calls them. Everything else stays out of context.

The numbers are not marginal. Anthropic shipped tool search and measured an 85 percent token reduction. Their code execution approach, where servers are exposed as code the model writes against, took one task from 150,000 tokens to 2,000. Cloudflare got that 1.17 million token API down to about 1,000.

If you want to apply this, order matters

Consolidate first. Most tool sprawl is not real. Thirty tools wrapping thirty endpoints of one API should be one tool with an operation parameter.

Then tier. Keep the handful you use every turn loaded upfront and defer the rest to names only. Then name things properly, because once tools are found by search, the name is the entire discovery surface. A badly named tool is an invisible tool.

What this means

The counterintuitive part is worth sitting with. An agent gets more capable when you show it less. Not because the tools are gone, but because choosing well requires being able to see the choice.

Sources

Anthropic, Advanced tool use / Tool Search Tool (Nov 2025); Anthropic, Code execution with MCP (Nov 2025); Cloudflare, Code Mode (Feb 2026).