Tool Calling

Introduction

Tool calling (a.k.a. function calling) is how agents take action. The LLM emits a structured JSON describing a tool + args; your code executes and returns the result.

Beginner analogy: a remote control — the LLM presses buttons, your code makes the TV change channels.

Understanding the topic

Core concepts:

• Tools are described by name + JSON schema.
• LLM picks tool + args; your runner executes.
• Always validate args (Zod, Pydantic) before running.
• Stream tool results back into context.
• Limit tool count (< 20) for accuracy.

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, MCP servers — all enable tool calling.

Best practices

• Write detailed tool descriptions; they're prompts.
• Validate args; assume LLM lies sometimes.

Common mistakes

• Granting tools without scopes — agent buys 100 widgets.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Steps in a tool call.
• Q2. Why validate args?
• Q3. Scenario: agent picks the wrong tool. Two diagnostic steps?