Tool use (function calling)

It is what lets a model check live data or change a record instead of relying on what it learned in training. AI agents are built on it.

How it works

You describe each tool to the model with a name, a short description and an input schema that lists the fields it accepts. When the model decides it needs a tool, it pauses and returns a request with the tool's name and the inputs. Your code runs the action and sends the result back, and the model continues. In Anthropic's API these messages are called tool_use and tool_result. Some tools, such as web search, run on the provider's side, so you only see the result.

The model reads tool descriptions the way it reads a prompt. Anthropic's guide to agents recommends giving tool definitions as much care as the main prompt, with example inputs and clear boundaries between similar tools.

Example

A freelance ops manager gives a model two tools: look up a contact in HubSpot, and create a contact. When a new form submission arrives, the model classifies the inquiry and calls the lookup. It calls "create" only if no match exists, and only after someone clicks approve in Slack.

Common mistake

Every tool on the list is something the model can call. If a tool can delete or send, a line in the prompt won't stop it. Give read tools freely, put an approval step in front of anything that writes, and keep the API keys in a credential store, never in the prompt.