Tool Calling

Introduction

Tool calling (a.k.a. function calling) lets the LLM choose which function to invoke and with what arguments. Provider APIs return the function name + args as structured JSON; your code runs the function and returns the result.

Beginner analogy: Like giving the AI a remote control — it picks which button to press; you actually press it.

Understanding the topic

Core concepts to understand:

• Define tools as JSON schemas (name, description, parameters).
• LLM decides which tool to call based on the user's request.
• Your code executes the tool, returns the result back to the LLM.
• Supported by OpenAI, Anthropic, Gemini, open models.

Syntax reference

Visual workflow / architecture:

User: "What's the weather in Tokyo?"
       │
       ▼
   LLM (sees tools: get_weather, search)
       │
       ▼  decides
   call get_weather(city="Tokyo")
       │
       ▼  your code runs it
   {"temp_c": 22, "rain": false}
       │
       ▼  back to LLM
   "It's 22°C and dry in Tokyo."

Real-world use

Powers ChatGPT plugins, Claude tool use, Cursor file actions, and every AI agent built since 2023.

Best practices

• Write clear tool descriptions — the LLM relies on them to choose correctly.
• Validate tool inputs (Zod / Pydantic) before executing.
• Always have a 'reply directly' option so the LLM doesn't force a tool.

Common mistakes

• Vague tool descriptions → wrong tool chosen.
• Forgetting to validate args → runtime crashes.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. How does tool calling work?
• Q2. How does the LLM choose which tool to call?
• Q3. Why validate tool args?