Creativity · Agent Protocol
Swarm Pattern — Peer Handoffs Between Agents
Swarm is a multi-agent pattern popularised by OpenAI's 2024 educational framework of the same name. Instead of a central orchestrator, each agent is a peer that can hand off the conversation to another agent using a tool call. Control jumps laterally between agents as the task evolves — more like a relay race than a supervisor-and-workers. Swarm later influenced OpenAI's production Agents SDK.
Protocol facts
- Sponsor
- OpenAI
- Status
- stable
- Spec
- https://github.com/openai/swarm
- Interop with
- OpenAI Agents SDK, function calling, A2A
Frequently asked questions
How is swarm different from orchestrator-worker?
Swarm has no lead agent. Any agent can hand off to any other by calling a handoff tool, and the conversation context travels with the baton. Orchestrator-worker keeps a single planner in charge.
When is swarm the right pattern?
When the task flows through stages that are best handled by specialists, and each specialist knows whom to hand off to next — for example triage → billing → refund — but no single brain needs to see the whole plan.
Is OpenAI's original Swarm framework still maintained?
The original swarm repo was explicitly educational. Its ideas are now carried forward in OpenAI's production-grade Agents SDK.
Sources
- OpenAI Swarm — educational repo — accessed 2026-04-20