OpenAI Agents SDK

What it is

The OpenAI Agents SDK is an open-source toolkit for building applications in which a language model does more than answer a single question: it calls tools, and it passes work on to other agents. OpenAI released the SDK in March 2025, in the same announcement that introduced the Responses API and built-in tools for web search, file search and computer use.^[1] The code is published in the openai/openai-agents-python repository on GitHub, which was created on 11 March 2025, is released under the MIT license, and has roughly 30,000 stars.^[2]

The problem it addresses is a practical one. Once a task needs several steps, needs to call outside tools, or needs several specialised agents to take turns, the scheduling logic in an application grows complicated — and that logic tends to get rewritten from scratch in every project.

Where it came from

The SDK is the successor to Swarm, an experimental SDK OpenAI released the previous year that was widely adopted by developers. OpenAI describes the Agents SDK as a significant improvement over Swarm, and frames its main purpose as simplifying the orchestration of multi-agent workflows — the plumbing that decides which agent acts when, and what it is allowed to do.^[1]

The four building blocks

OpenAI lists four improvements that shape the design of the SDK.^[1]

Agents are language models with instructions and built-in tools, easily configured. The agent is the unit that reads a task and decides what to do next.

Handoffs transfer control between agents intelligently, so a request that starts with one agent can finish with another that is better suited to it.

Guardrails are configurable safety checks for input and output validation, letting developers set limits on what an agent may accept and what it may return.

Tracing and observability leave a record of what happened during a run, so that a multi-step, multi-agent process can be inspected afterwards instead of being treated as a black box.

Which models it works with

The SDK works with the Responses API and the Chat Completions API. It also works with models from other providers, as long as they provide a Chat Completions style API endpoint.^[1] At launch the SDK was Python-first, with Node.js support announced as coming soon.^[1]

Early users

OpenAI cited two companies at launch. Coinbase used the SDK to prototype and deploy AgentKit in a few hours. Box created agents with web search in a couple of days.^[1]

Where it sits

The SDK arrived as part of a broader set of agent-building tools: the Responses API and the built-in tools shipped in the same announcement.^[1] It is open source under a permissive license, and its repository carries a star count in the tens of thousands.^[2] What it standardises is the scaffolding around the model — how an agent is defined, how work is routed between agents, what is checked on the way in and out, and how a run is recorded.