OpenAI Swarm

Experimental pattern for lightweight agent "Handoffs".

1. Philosophy

Swarm is stateless. An agent is just a set of instructions. To "Orchestrate", Agent A simply calls a function that returns Agent B. This is called a Handoff.

Technique: The Handoff

from swarm import Swarm, Agent

client = Swarm()

# 1. Define Handoff Logic
def transfer_to_spanish():
    """Transfer spanish users here."""
    return spanish_agent

# 2. Define Agents
english_agent = Agent(
    name="English",
    instructions="You speak English.",
    functions=[transfer_to_spanish] # <--- The Handoff Tool
)

spanish_agent = Agent(
    name="Spanish",
    instructions="Tu hablas español.",
)

# 3. Run
# If user says "Hola", agent calls 'transfer_to_spanish', and Swarm switches active agent.
response = client.run(
    agent=english_agent,
    messages=[{"role": "user", "content": "Hola"}]
)
print(response.messages[-1]['content'])

Technique: Context Variables

Since Swarm is stateless, you pass data manually.

def greet(context_variables):
    name = context_variables.get("name", "User")
    return f"Hello {name}"

agent = Agent(functions=[greet])

client.run(
    agent=agent,
    messages=[...],
    context_variables={"name": "Alice"}
)