Tool Use — AI Glossary

Reference details

Related terms

• Function Calling
• AI Agent
• MCP (Model Context Protocol)
• Code Interpreter

How it works

1. A tool reaches the model as a name, a description, and a typed parameter list, so the description is the entire basis on which the model decides whether the tool fits.
2. That makes the description an interface specification rather than documentation: a vague one produces wrong tool selection in a way no amount of model capability compensates for.
3. The result must come back as text the model can read, which means a tool returning a status code or an opaque handle has to be wrapped before it is useful.
4. Errors are returned as content rather than raised, because the model can respond to a traceback it can see and cannot respond to an exception the harness swallowed.
5. The model observes only what the result says, so a side effect that succeeded silently is indistinguishable to it from one that never ran — which is why confirmations have to be in the returned payload.
6. Nothing in the mechanism authenticates or bounds anything: the model produces a request to act, and every permission, rate limit and approval sits in the harness that decides whether to honour it.

Example: Where the model stops being able to guess

Ask a model for the 14th root of 892,371 and it will produce a confident, usually wrong number, because it is predicting digits. Give it a calculator tool and it emits pow(892371, 1/14), gets 2.534…, and reports that. The model did not get better at arithmetic; it stopped doing arithmetic.

Commonly confused with

Tool use is the capability; function calling is the interface that exposes it. A model can use tools through several mechanisms — a tool-calling API, MCP, or plain code generation that your runtime executes — and function calling is only the most common one.

When to use it

Reach for it when:

• Exactness matters and the model is unreliable: arithmetic, dates, lookups
• The information postdates training, or is private to your tenant
• The task has a real-world effect — sending, writing, booking

Reach for something else when:

• The model already knows the answer well; a tool adds latency and a failure mode
• You cannot authenticate or rate-limit the tool being called
• The tool's output is untrusted text that flows straight back into the prompt — that is the prompt-injection path

Referenced by

This term is used by the following ModelRefs references:

• function calling prompt
• react
• react
• plan and execute
• single tool agent
• multi tool router
• code execution agent
• learn agents
• learn mcp
• master langgraph
• anthropic builder path

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Tool Use — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path