OpenAI Introduces Swarm, a Framework for Building Multi-Agent Systems

As the AI world is shifting towards building agentic systems, OpenAI’s team has decided to give the world an open source gift. The team has released Swarm, an open source framework for building, orchestrating and deploying multi-agent systems.

Click here to check out the GitHub repository. 

Managed by the OpenAI Solutions team, the framework is still in its experimental phase and is currently not intended for production.

Swarm is designed to make agent coordination and execution lightweight, highly controllable, and easy to test. It achieves this using two core abstractions: Agents and handoffs. An Agent holds instructions and tools and can transfer control to another Agent at any time.

These simple yet powerful abstractions enable rich interactions between tools and networks of agents, allowing for scalable, real-world solutions without a steep learning curve.

Swarm Agents are not the same as Assistants in the Assistants API. Though similarly named for convenience, they are entirely different. Swarm operates with the Chat Completions API and remains stateless between calls.

While the Assistants API offers a fully-hosted solution with built-in memory management, Swarm is perfect for developers seeking full transparency and granular control over context, steps, and tool usage. Running almost entirely on the client, Swarm—like the Chat Completions API—does not retain state between calls.

Swarm is lightweight, scalable, and highly customisable, ideal for scenarios with many independent capabilities and instructions that don’t fit into a single prompt.

Along with this, OpenAI also released MLE-bench, a benchmark designed to assess how effectively AI agents can perform machine learning engineering tasks. To create this, the team compiled 75 ML engineering-related competitions from Kaggle, offering a wide range of challenges that test key skills like model training, dataset preparation, and experiment execution. 

OpenAI evaluated several language models on this benchmark using open-source agent scaffolds. Notably, the best-performing configuration — OpenAI's o1-preview with AIDE scaffolding — achieved the level of a Kaggle bronze medal in 16.9% of the competitions.