Tool Calling Basics

Introduction

Tool calling lets the LLM emit a structured request to invoke a function — your code runs it and returns the result. This is how agents act in the world.

Beginner analogy: a manager (LLM) tells the assistant (your code) 'book the flight'. The assistant uses the actual booking system, then reports back.

Understanding the topic

Core concepts:

• Each tool has a name, description and JSON schema.
• LLM picks a tool + args; your code executes it.
• Tool result feeds back into context for next decision.
• Always validate args before execution.
• Use few-shot tool examples for tricky ones.

Syntax reference

Visual workflow / architecture:

LLM decides
    │
    ▼
{ tool_name, args }
    │
    ▼
┌──────────────┐
│   Validator  │  schema check
└──────┬───────┘
       ▼
┌──────────────┐
│  Tool runner │──► HTTP · DB · Shell
└──────┬───────┘
       │ result
       ▼
   Back to LLM

Real-world use

OpenAI function calling, Anthropic tool use, Gemini functions, every agent framework — all expose this pattern.

Best practices

• Keep tool names short and descriptive.
• Write detailed tool descriptions — they're prompts!
• Limit to < 20 tools per agent.

Common mistakes

• Vague tool descriptions → wrong tool chosen.
• No arg validation → SQL injection, money loss.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What does the LLM actually return for a tool call?
• Q2. Why validate args before execution?
• Q3. Scenario: agent calls the wrong tool. Two fixes?