OpenAI shipped an experimental framework it warned against using

The conversation

In October 2024, OpenAI published Swarm on GitHub, a repository whose own README names it "Swarm (experimental, educational)" and states it "explores patterns that are lightweight, scalable, and highly controllable" for coordinating more than one agent. The earliest capture of the repository the Internet Archive holds, from 23 October 2024, already carries that framing: an archived snapshot shows the same "experimental, educational" label and no reference to any successor product, since none existed yet. The README states Swarm runs through two primitives, Agents and handoffs, where an agent holding a conversation can at any point pass it to another agent.

What the documents show

The README is explicit that Swarm is not related to the Assistants API already on file: "Swarm Agents are not related to Assistants in the Assistants API," and Swarm instead runs "entirely on the client," powered by the Chat Completions API, storing no state between calls. OpenAI released the code under the MIT License, permitting reuse without a separate agreement, and the repository's own listed examples include a triage agent, an airline customer-service setup and a personal-shopping agent, described as illustrations rather than production templates.

The system boundary

The README draws its own boundary against production use throughout: it calls Swarm "an educational resource for developers curious to learn about multi-agent orchestration," contrasted with the Assistants API, which it says is "a great option for developers looking for fully-hosted threads and built in memory management and retrieval." That is OpenAI stating, in its own words, which of its two products handles state and infrastructure and which does not, rather than leaving a builder to infer it.

Where it fails

As retrieved 16 September 2026, the same repository's README now opens with a notice absent from the October 2024 version: "Swarm is now replaced by the OpenAI Agents SDK, which is a production-ready evolution of Swarm," recommending migration "to the Agents SDK for all production use cases." OpenAI's own GitHub listing for the repository still describes it as "managed by OpenAI Solution team," and the repository has not been archived or deleted, so its code stays runnable even though the same document now tells a reader not to use it that way.

• Does a repository's current README still carry the caution label its first version did, or has the guidance changed?
• Is an experimental framework's stated successor a like-for-like replacement, or a different API surface entirely?
• Does a project's own documentation say who maintains it day to day, and has that changed since launch?

Swarm's own documents describe a project that named its own limits before anyone else did, then pointed away from itself once a production-ready successor existed, a sequence this desk can trace because OpenAI wrote each stage down in the same file.