OpenAI Swarm
A lightweight, experimental framework by OpenAI for building multi-agent systems with minimal abstraction. Focuses on explicit agent handoffs and stateless design.
Overview
Swarm is an open-source experimental framework released by OpenAI that provides minimal, ergonomic primitives for building multi-agent systems. It is intentionally lightweight — no state persistence, no complex graph abstractions — designed for teams who want explicit control without a heavy framework dependency.
Core Concepts
- Agents — defined by a system prompt, a model, and a list of available functions (tools).
- Handoffs — an agent can transfer control to another agent by returning a reference to it. The receiving agent takes over with full context.
- Context variables — a dictionary of shared state passed between agents across handoffs.
When to Use Swarm
Swarm is best for:
- Simple multi-agent pipelines with a small number of well-defined agent roles.
- Prototyping and learning — the minimal API makes it easy to understand what is happening.
- Teams who want to understand agent coordination at the primitive level before adopting a heavier framework.
It is not suitable for production systems requiring persistence, complex branching, or observability tooling.
Key Limitation
Swarm is stateless by design. Each .run() call is independent. Long-running workflows that need to resume across sessions require building your own state persistence layer on top.
Resources
- GitHub:
openai/swarm - Experimental status: not intended for production use without modification.