[Day 5/100] Tool Use and Function Calling, Explained from Scratch

2026-06-21 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Novice, medium

Summary

The article explains that AI model "function calling" is not actual function execution but rather a convention layered on text generation. It details how the OpenAI API serializes tool schemas into a special prompt section, biases the model to generate structured output, and then parses this output into `tool_calls`. The full lifecycle involves user input, building messages with tool schemas, model response generation, API parsing, user code execution, appending results, and the model generating a final answer. Effective tool design requires clear, short verb names, comprehensive descriptions that guide model usage, and precise parameter schemas including types, examples, and `enum`s to ensure consistent argument values. It also highlights the importance of specific error handling within tools.

Key takeaway

For AI Engineers building agentic systems, understanding that function calling is text generation, not direct execution, is crucial. This insight helps you debug malformed calls and design robust tools. You should prioritize clear tool names, comprehensive descriptions, and precise parameter schemas with `enum`s to guide model behavior. Additionally, implement specific, recoverable error handling within tools to improve agent recovery and reliability.

Key insights

Function calling is trained, constrained text generation parsed into structured data, not direct function execution.

Principles

• Tool descriptions dictate model selection between competing tools.
• Parameter `enum`s ensure consistent argument values.
• Design tools for independence to enable parallel execution.

Method

The function calling lifecycle involves user input, API interaction for model response and parsing `tool_calls`, user code execution, appending results, and a final model response to the user.

In practice

• Use short, clear verb names for tools like `get_weather`.
• Return specific error messages as tool results, not exceptions.
• Combine related functions into single tools like `get_user_profile`.

Topics

• Function Calling
• Tool Use
• Agentic AI
• LLM Prompt Engineering
• OpenAI API
• Error Handling

Best for: AI Engineer, Machine Learning Engineer, AI Student

Related on AIssential

• Day 14: Function Calling Explained Simply (For DevOps & Cloud Engineers) — Function Calling empowers LLMs to interact with external systems, transforming them from text generators into actionable AI agents.
• Tool Calling, Explained: How AI Agents Decide What to Do Next — Tool calling enables LLMs to trigger external actions by deciding which function to call and with what arguments, forming agentic AI workflows.
• Gemma 4 Tool Calling Explained: Build AI Agents with Function Calling (Step-by-Step Guide) — Gemma 4's native tool calling enables local AI agents to reliably interact with external APIs for real-world data.
• Structured Outputs vs. Function Calling: Which Should Your Agent Use? — Structured outputs define data shape; function calling dictates application control flow and external interactions.
• Day 15: Tool Usage in AI Agents (For DevOps & Cloud Engineers) — Tools transform AI agents from chatbots into active assistants by enabling real-world system interaction.
• To Call or Not to Call: A Framework to Assess and Optimize LLM Tool Calling — LLM tool-calling decisions, especially for web search, often misalign with true necessity and utility.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.