OpenAI Swarm — Lightweight Multi-Agent Orchestration

Core Concepts

Agents

Agents are instructions + functions. Simple and composable:

agent = Agent(
    name="Support",
    instructions="You handle customer support.",
    functions=[lookup_order, check_status, escalate],
)

Handoffs

Agents transfer control by returning another agent:

def transfer_to_billing():
    return billing_agent  # Control passes to billing

Context Variables

Share state across agents without global variables:

response = client.run(agent=triage,
    messages=[...],
    context_variables={"user_id": "123", "plan": "premium"})

Routines

Multi-step procedures encoded as agent instructions — like a playbook for the agent to follow.

FAQ

Q: What is OpenAI Swarm? A: An educational framework by OpenAI exploring lightweight multi-agent orchestration with handoffs, tool calling, and context variables. 21K+ stars, MIT licensed.

Q: Should I use Swarm in production? A: OpenAI explicitly states Swarm is educational, not a production framework. For production, consider OpenAI Agents SDK, LangGraph, or CrewAI.