tool search

advancedspreading

An AI agent does not retain descriptions of all its tools in memory but retrieves and loads only the few it needs during a task.

In plain terms

Imagine a mechanic who does not haul the entire workshop around. The car holds the warehouse's inventory list; if a pipe wrench is needed, the mechanic looks it up and brings in only that. Tool search enables the same for an AI agent.

An agent today can access hundreds of tools: email, calendar, file manager, database, corporate systems. Each comes with a detailed description of its purpose and how to use it. If all of these are presented to the model at once, its working memory (the context window) fills up before it even starts the actual task.

With tool search, the agent initially sees only the tool names or brief descriptions. It retrieves the full user manual only when truly needed. This allows it to work more cheaply and quickly, and choose the wrong tool less often, as it does not have to select from hundreds of similar options.

How it works

The developer marks less frequently used tools as "lazy-loaded": their detailed descriptions are not pre-loaded into the model's working memory, only their names.
The agent receives a dedicated search tool to query the tool catalog by keywords, name, or free-text description.
When a task requires a capability not immediately available, the agent formulates a search, such as 'sending an email' or 'calendar event'.
The system returns the full description of the top few matches: what the tool does and what data it expects.
From now on, the agent can call the found tool as if it had been there from the start, while the others continue to take up no space.

An everyday example

You ask your AI assistant to add next week's dentist appointment to your calendar. The assistant has access to your email, documents, and calendar, but it only looks up and loads the calendar tool and does not touch the dozen or so others. All you see is that it responds quickly and does not get stuck even in longer conversations.

Who is affected

developerIf you connect many tools or multiple MCP servers to your agent, without tool search the descriptions can consume a significant portion of the context window, reducing tool selection accuracy. → Code

entrepreneur, executiveThis determines how many internal systems a corporate agent can connect to simultaneously without usage costs and response times skyrocketing.

employeeThis is why your AI assistant can work with many applications simultaneously, and why it may sometimes fail to find an existing capability. → Work & HR

A common misconception

MythTool search means that the AI searches the internet for new programs and installs them.

In realityThe agent only searches within the catalog that the developer or user has pre-assembled and authorized for it. It does not acquire new capabilities; it simply selects the necessary ones from the existing set.

Limits and risks

If a tool's name or description is inaccurate, the search will not find it, and the agent may behave as if that capability does not exist.
Search is an extra step: unnecessary slowdown for few tools, truly worthwhile only with dozens or hundreds of tools.
Search is not access control: just because an asset is findable, rules must still be set separately on what the agent can do with it.
Implementation varies by provider, so the same toolset may behave differently in different systems.

Similar, but not the same

tool use – Tool use is the action itself when an agent invokes a tool, while tool search is the preceding step in which it selects which one to load.
web search – Web search looks for information on the internet, while tool search looks for capabilities within the agent's own tool catalog.
retrieval-augmented generation – Retrieval-augmented generation (RAG) retrieves document excerpts for the content of an answer, while tool search retrieves tool descriptions for taking action.

Related concepts

Broader

tool use, kontextusmenedzsment, AI-agent

Narrower

deferred asset loading

On the edge

Model Context Protocol, context window, function calling, agent skills, retrieval-augmented generation

A brief history

The need became urgent in November 2024, when the Model Context Protocol (MCP) standard published by Anthropic led to agents being connected to tools en masse. In November 2025, Anthropic introduced its Tool Search Tool solution on its developer platform, initially in beta.

Where things stand

By late 2026, it is an increasingly established solution in agent systems that work with many tools and in coding agents. The user typically sees nothing of it, as it works in the background.

Status: 2026 November

Also known as

tool search · tool search · Tool Search Tool · tool search · On-demand tool loading · deferred asset loading · deferred tool loading · tool discovery

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.