tool use

On its own, a language model can only produce text — it cannot actually check today's weather, do exact arithmetic, or send an email. Tool use gives it hands: a set of external functions it is allowed to call, like 'search the web', 'run this calculation', or 'look up this customer'. The model writes out which tool it wants and with what inputs, the program runs it, and the real-world result comes back into the conversation.

The model never runs the tool itself. You describe each available tool — its name, what it does, and the inputs it expects — and the model learns to emit a request matching that description when it judges the tool would help. The surrounding system executes the request and feeds the output back as new context the model reads on its next turn. Tools turn a closed text generator into something that can observe and affect the outside world.

This is what makes models accurate and current rather than guessing from memory. But every tool is also a new surface for things to go wrong — wrong inputs, failed calls, or actions taken too eagerly — so tools are usually granted carefully, with validation and limits on what they can touch.