OpenAI Swarm

Swarm is OpenAI's teaching version of multi-agent orchestration: an MIT-licensed repository from its Solutions team, still headed 'experimental, educational' in the README. An agent is instructions plus Python functions, and only two operations exist — a routine, which loops over tool calls until the agent answers, and a handoff, where one agent returns another to carry on. There is no graph, no state store, no memory and no scheduler; the loop is plain code over the chat-completions interface, readable in an afternoon. That is why it mattered: it named the handoff pattern larger frameworks had buried in their machinery.

Judgement: on 2026-09-20 the repository is public, unarchived but effectively frozen — newest commit 15 April 2026, the one before it March 2025 — and the README states plainly that Swarm 'is now replaced by the OpenAI Agents SDK', recommending migration for production use. Maintenance went to the successor, which published v0.22.3 on 17 September 2026 and was still taking commits the day this was checked. It installs from git over SSH, so there is no version to pin, no changelog and no upgrade path. Nothing needs hosting: it is client code, so an API key is the only requirement, and it targets the older chat-completions shape, not the responses APIs current SDKs use. No streaming controls, guardrails, tracing or sessions ship with it, and it loses anywhere a run must survive a restart. Read it, teach handoffs from it, then build on the Agents SDK.