With tool calling, you describe the tools an AI model may use: a name, what it does and what inputs it takes. When the model decides a tool is needed, it does not run it. It returns a structured request naming the tool and the arguments. Your code runs the real action and sends the result back, and the model continues with that result in view.
This is how a large language model gets beyond text: it can look up live data, write to a system or trigger a workflow. Because your code executes the action, you control what is allowed, and you should validate the arguments before acting, since the model can produce wrong or unexpected values.
Tool calling and function calling refer to essentially the same mechanism. Tools are what make an AI agent possible.
Related: API, Error Handling, Idempotency.