AI Trend Research

← AI glossary

AI glossary · Basic · Agents

Tool use

Tool use is a capability that lets an AI model call outside tools such as search, calculators, databases and external APIs instead of producing every answer on its own. It is also called function calling.

About this term
Level
Basic · Technical terms you often meet in AI news
Fields
Agents
Added
2026-10-08
In plain words It is like an assistant who, instead of memorizing everything, contacts the right person and hands off work when needed. The assistant only decides whom to ask for what, and the person in charge does the actual work and reports back the result.

How it works

The developer describes in advance the name, purpose and required inputs of each tool the model can use. When the model receives a question and decides a tool is needed, it outputs a request in a set format (usually JSON) saying 'run this tool with these values'. The model is not the one that runs the tool. The app or server that receives the request actually runs the tool and returns the result to the model. The model reads the result and either finishes the answer or repeats the process by calling another tool.

The tool-use loopUser question → LLM; LLM → Tool call request(If needed); Tool call request → Tool execution; Tool execution → Execution result; Execution result → LLM(Return result); LLM → Final answer(If enough)User questionLLMDecide whether a toolis neededTool call requestTool name + inputsTool executionApp/server runs it(search, calculation,API)Execution resultFinal answerIf neededReturn resultIf enough
The tool-use loop When the model needs a tool, it only outputs a call request; the app or server does the actual execution. The result goes back to the model, which decides again and gives a final answer once it has enough.

Why it matters

LLMs do not know about events after their training cutoff and are weak at exact calculation and at looking up internal company data. With tools, they can search for current information, calculate accurately and even take real actions such as sending email or adding calendar events. AI agents, coding agents and computer use features are all built on this capability. Standards such as MCP emerged to make it easy to connect one tool to many AI apps.

Things to know

Failures are common: the model picks the wrong tool, passes wrong input values or misinterprets the result. For tools that take hard-to-reverse actions such as payments or deletions, it is safer to add a step where a person confirms before execution. Results a tool brings back, such as web documents, may hide malicious instructions, so there is also a risk of prompt injection. The more tools there are, the more context the tool descriptions take up, and the harder it becomes to pick the right tool.

Example

A common example of tool use is a chatbot that receives a question, runs a web search and answers based on the results. When a coding agent reads files, runs commands and runs tests, all of it is done through tool calls. The paper covered on this site, 'From Evidence to Action: Where Do Tool-Using Agents Fail?', also looks at the steps where such agents fail.

Where it came up on this site

Related terms

AI agent MCP Computer use Coding agent Prompt injection Context engineering