Tool Use
Letting a model call external tools, APIs, or services as part of solving a task.
Where did the opening time come from?
A visitor asks when a fictional archive’s reading room opens on Saturday. The assistant has access to a calendar maintained by the archive.
- The request
Ask the calendar for Saturday hours at the reading room. The room and date distinguish this lookup from the main building’s schedule.
- The returned record
The calendar reports a special opening at 11 a.m. for that Saturday. The service also returns the date and location.
- The answer
The assistant gives 11 a.m. and identifies the calendar entry. If the lookup fails, the current opening time is still unknown.
What did the tool contribute?
The calendar supplied a fact that can change. The model interpreted the question and turned the result into a sentence. Recording the returned date and room lets someone check a plausible but misplaced answer.
Why Tool Use Needs Control
Giving a model tools introduces risk as well as power. A tool may expose sensitive data, trigger external effects, or produce output that should not be trusted automatically. That is why tool-enabled systems usually rely on guardrails, explicit permissions, structured tool schemas, and confirmation steps for high-impact actions.
The best way to think about tool use is not as "letting the model do anything," but as designing a safe and limited workspace in which the model can help. Strong tool systems define what the model can call, what arguments are allowed, and what happens if a tool fails or returns bad data.
Follow the idea
Function Calling · Inspect the arguments an application validates.
AI Agent · See how an observation can guide a subsequent action.
Related concepts
Examples in the AI guides
Related Yenra articles (2)
Source · Anthropic on designing tools for agents
Return to Tool Use in the A–Z glossary