A model can propose using a tool: an operation provided by the surrounding software. The application handles the request, applies its permissions and returns a result. Tools might search an approved collection, calculate a value, read a file or send a message.

The distinction matters: the model’s proposed call is not proof that anything happened. Check the actual tool result and, for an important change, the resulting external state.

A tool also has limits. A search can miss a source; a calculation can use the wrong inputs; an account can lack permission. Keep those limits visible when judging the final answer.