Towards Data Science explains how large language models move beyond static replies by using tool calling, a mechanism that lets an AI select and invoke external functions. The model analyzes the user prompt, matches it to a declared function calling schema, and returns a structured call instead of plain text. Execution happens in user code, after which the result feeds back to the model.

The article walks through a concrete example: a weather assistant built on the free Open-Meteo API. After defining a get_current_weather function with city and unit parameters, the model receives a query like “What’s the weather in Athens?” and responds with a tool call containing those arguments. The developer then runs the API request, captures temperature data, and returns it to the LLM for a final answer.

Tool calling separates decision‑making from execution, preventing the model from directly accessing external resources while still enabling real‑time actions such as database queries, messaging, or API calls. Clear function descriptions and strict JSON schemas are essential for the model to choose the correct tool. This pattern underpins emerging agentic AI workflows that can act autonomously within defined safety bounds.