// the find
openai/swarm
Educational framework exploring ergonomic, lightweight multi-agent orchestration. Managed by OpenAI Solution team.
Swarm is OpenAI's lightweight, stateless framework for multi-agent orchestration built around two primitives: Agent (instructions + functions) and handoffs (returning another Agent from a function call). It's explicitly labeled educational/experimental and has been superseded by the openai-agents-python SDK for anything production.
The Agent+handoff model is genuinely minimal and easy to reason about — an agent is just instructions plus a function list, and handing off is nothing more than returning a different Agent object from a tool call. Function-to-JSON-schema conversion from type hints and docstrings removes a lot of Chat Completions boilerplate. The examples directory (airline, triage_agent, support_bot) covers real multi-agent patterns like routing and context-variable passing, not just toy demos. Streaming is supported with explicit delim events so you can detect agent switches mid-stream.
The README leads with a deprecation notice pointing to openai-agents-python — this repo is frozen in an educational, non-maintained state, so anyone starting fresh should go there instead. It's stateless between calls by design, meaning you own all persistence, conversation history, and context storage yourself. It's hard-coupled to the OpenAI Chat Completions API with no abstraction for other providers. Error handling is minimal: a failed function call gets appended as an error message for the agent to 'recover gracefully,' but there's no retry, validation, or guardrail layer built in.