# What is Tool Use?

Also called tool calling, agent tools.

Tool use is an agent's ability to invoke external functions such as search, database queries, file operations, or third-party APIs, and to incorporate the results into its next decision. The model does not run the tool itself. It emits a structured request naming the tool and its arguments, and the surrounding runtime executes it and returns the output.

Each tool is described to the model by a name, a purpose, and a schema for its arguments. The model chooses one and produces arguments matching the schema. The runtime validates the request, checks that the agent is permitted to make it, executes it, and appends the result to the conversation. From the model's perspective a tool result is just more context to read.

Tools are what connect a model to anything outside its training data and its context window. They supply current information, authoritative records, exact computation, and the ability to change something in the world. Without tools an agent can only produce text about what it already knows. With them, its reliability becomes a function of the tools' quality as much as the model's.

Tool design is underrated. Overlapping tools cause wrong selections, vague descriptions cause wrong arguments, and enormous result payloads flood the context window and crowd out the actual task. Too many tools at once degrades choice accuracy. The usual remedies are fewer and sharper tools, descriptions written for the model rather than for developers, and truncated or summarized outputs.

Every tool is also an authority boundary. A tool that can send, delete, or spend turns a model mistake or a malicious instruction hidden in retrieved content into a real consequence, so permissions, approvals, and audit belong at the tool layer rather than in prompt wording. Related terms are function calling, which is the mechanism, and the agent loop, where calls happen.

## Key points

- The model requests a tool; the runtime executes it.
- Tools supply current data, exact computation, and real effects.
- Vague descriptions and overlapping tools cause most selection errors.
- Each tool is a permission boundary and needs its own controls.

## In practice

Asked how many orders shipped yesterday, a model with no tools can only guess. Given a query tool, it emits a request naming the tool and a date range, the runtime runs the query against the real database and returns eighty-seven rows, and the agent answers with the exact figure. The number came from the database. The model only decided what to ask for.

## Related terms

- [Function Calling](/en/glossary/function-calling)
- [Agent Loop](/en/glossary/agent-loop)
- [AI Agent](/en/glossary/ai-agent)
- [Agent Architecture](/en/glossary/agent-architecture)
- [Agent Skill](/en/glossary/agent-skill)

[Back to the AI Glossary](/en/glossary)