Swarm

Swarm is an experimental, educational Python framework from OpenAI for exploring lightweight multi-agent orchestration. It is built around two primitives, Agents and handoffs, and runs on top of the Chat Completions API. The README states that Swarm has been replaced by the OpenAI Agents SDK, which it recommends for production use cases.

What It Is

Swarm lets developers define Agents, each with instructions and a list of Python functions, and let an Agent hand a conversation to another Agent. A Swarm client wraps an OpenAI client, and its run method takes messages and returns messages, the last active Agent, and updated context variables. It is stateless between calls and runs almost entirely on the client.

How the Execution Loop Works

The run loop gets a completion from the current Agent, executes tool calls and appends results, switches Agent if a function returned one, updates context variables, and returns when there are no new function calls. Functions are automatically converted to JSON Schema for Chat Completions tools, using docstrings and type hints. Instructions can be a string or a function that receives context variables. Streaming is supported, with added start/end delimiter events and a final aggregated response.

Examples and Testing

The repository includes examples such as basic setup, a triage agent, a weather agent, an airline customer service setup, a support bot, and a personal shopper. A REPL demo loop allows testing agents from the command line, and the authors encourage developers to bring their own evaluation suites.

Current Status

The README marks Swarm as experimental and educational, and states it is replaced by the OpenAI Agents SDK, which the OpenAI team will actively maintain. It requires Python 3.10+ and is installed from the GitHub repository under the MIT License.