Tool Calling Concepts

Introduction

Tool calling (a.k.a. function calling) lets the LLM emit a structured request to call your code. You run the code, return the result, and the model continues.

Beginner analogy: Like a manager delegating: 'go fetch X from the API'. The model decides what to delegate; your code does it.

Understanding the topic

Core concepts to understand:

• Define tools with JSON schema (name, description, parameters).
• Model emits a structured 'tool call' instead of text.
• Your code executes the tool and feeds the result back.
• Modern APIs (OpenAI, Anthropic) have first-class support.

Syntax reference

Visual workflow / architecture:

tools = [
  { name: "get_weather",
    parameters: { city: "string" } },
  { name: "send_email",
    parameters: { to, subject, body } }
]

Model emits: { tool: "get_weather", args: { city: "NYC" } }
Code runs it → returns 22°C → model continues.

Real-world use

Every modern AI agent uses tool calling for browsing, code execution, database queries, sending emails.

Best practices

• Describe each tool with WHY to call it, not just WHAT it does.
• Cap to ~5 tools per prompt — more confuses the model.
• Validate tool arguments before executing.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What is tool calling?
• Q2. Difference from prompt-engineered 'fake' tools?
• Q3. How handle a wrong tool call?
• Q4. Why limit number of tools?