Products·OpenAI·Jun 2023

39. Function Calling and the Agent Ecosystem

Enabling language models to use tools through structured function calls

Product Announcement

Summary

OpenAI introduced function calling in the Chat Completions API, allowing GPT-4 and GPT-3.5-turbo to generate structured JSON arguments for developer-defined functions. This API feature enabled the entire AI agent ecosystem: LangChain, AutoGPT, and hundreds of tool-using agent frameworks were built on function calling.

Key Concepts

Structured tool use: JSON Schema function definitions in the API

Developers could define functions with names, descriptions, and parameter schemas. The model would decide when to call a function and generate valid JSON arguments. This turned language models from text generators into decision-making agents that could interact with external systems — query databases, call APIs, execute code, and chain multiple tool calls together.

Enabling the agent ecosystem: from text completion to tool orchestration

Function calling was the technical enabler for the AI agent wave of 2023-2024. LangChain, AutoGPT, BabyAGI, CrewAI, and dozens of agent frameworks were built on the ability of models to reliably call tools. Without function calling, "AI agents" would have remained unreliable prompt-engineering hacks. With it, they became a legitimate application pattern.

Parallel function calling and forced function use

Later updates added parallel function calling (multiple tool calls in one response) and the ability to force the model to call a specific function. These refinements made agent architectures more efficient and predictable, enabling complex multi-step workflows where the model could query multiple data sources simultaneously.