OpenAI Agents SDK: OpenAI's official agent framework

A deliberately small toolkit

OpenAI has published the Agents SDK since March 2025, first for Python and then for TypeScript, under the MIT license. The documentation frames it as the grown-up version of Swarm, an earlier experimental project. The goal is to give you just enough building blocks for a serious AI agent without forcing a heavy architecture on you.

Three ideas sit at the core:

• Agent: a large language model with instructions and tools, meaning functions it can trigger through tool calling.
• Handoff: one agent passes control to a more specialized one, which is how you build a multi-agent system.
• Guardrails: checks on inputs and outputs that can stop an off-topic request or a non-compliant answer.

Around these, the SDK adds sessions (conversation memory), human approval before sensitive actions, built-in tracing to inspect every step, connections to MCP servers, voice agents and “sandbox agents” that work inside a container on long-running tasks.

What you can build with it

Typical projects include a customer support desk where a triage agent routes people to a “billing” or “tech” agent, an internal assistant that queries a database and writes a summary, or a voice agent that answers phone calls. The SDK runs the agent loop for you (call the model, run the tool, call again); you focus on the business logic.

Getting started

You need Python 3.10 or later. Install the openai-agents package with pip, put your API key in an environment variable, define an agent in a few lines and run it with the SDK’s “Runner”. The official docs include ready-to-adapt examples, including for the TypeScript version.

Strengths and limits

Strengths: few concepts to learn, clear documentation, tracing included, permissive license, and support for other model providers.

Limits: the SDK is still designed first for OpenAI’s ecosystem, and some features (realtime, hosted tools) only work with its models. It doesn’t offer a state graph as detailed as some other frameworks, and there is no turnkey hosting: you deploy it yourself. Every model call also costs money; the AI agent cost calculator helps you estimate it.

Alternatives

If you need fine-grained control over each step, LangGraph is a better match. Google ADK supports more programming languages, Pydantic AI focuses on strict typing, and the Claude Agent SDK gives you an agent that already comes with tools. Browse the full list of agent frameworks to compare.