Tool search (deferred tool loading)

  • Agents
  • Industry term
3 steps

Tool search lets an AI agent find the actions it needs and receive instructions for using them. An agent is an AI app that can take steps such as searching files or sending messages.

A tool is an action the app lets the model take. The model is the trained AI that generates replies and chooses actions. Before using a tool, it needs instructions describing the action and the information to supply. Receiving those instructions only when needed leaves more room for your messages.

Where you’ll meet it

You've given your agent access to email, files, team chat, customer records, and a calendar. A warning says instructions for these actions use up much of the text the model can receive at once. Most of your tasks use only a few connections.

What to do

Look for an app setting that sends tool instructions to the model only when needed. If someone is building the agent for you, ask them to keep everyday tools ready and find others through search. Clear names and descriptions help the agent find the action you need.

Easy to mix up

Which problem does it address?
Tool search (deferred tool loading)Model Context Protocol (MCP)
Finding which connected tools' instructions to send to the modelGiving AI apps a shared way to connect to tools and data

Check yourself

Your agent searches the available tools for a way to find overdue invoices. The search succeeds. What should you expect it to receive before it can produce the invoice list?

Go deeperSearch can load detailed instructions while leaving earlier text unchanged

Anthropic introduced its “Tool Search Tool” in November 2025. Providers differ in how much they load in advance. OpenAI's tool-search documentation says individual actions can keep their names and short descriptions visible, while the detailed instructions for carrying out those actions wait to be loaded.

Anthropic and OpenAI document adding newly found tool definitions at the end of the conversation. This can preserve prompt caching, which reuses work from processing an identical beginning, because the earlier text stays unchanged.

Sources

Practice Modern AI

Or take the learning quiz on a mix of every set