The developer describes available tools; when the model decides one is needed, it emits a structured call ("run_tests with these arguments"), the application executes it, and the result goes back into the context for the model to read. Text-only models talk about doing things; tool use is how they actually do them.

It's the foundation agents stand on, and its failure modes are where agent debugging lives: wrong tool, wrong arguments, misread results, or looping on a failing call.