OpenAI Swarm

OpenAI Swarm is an experimental, educational framework for orchestrating multiple agents. It prioritizes a lightweight, stateless approach to multi-agent systems, focusing on Handoffs and Routines.

Core Primitives

1. The Agent

An Agent encapsulates:

• Instructions: The system prompt (can be a string or a callable returning a string).
• Functions: A list of python functions the agent can call.
• Model: The underlying LLM (defaults to gpt-4o).

2. Handoffs

A handoff occurs when an agent's function returns another Agent object.

• Mechanism: When the framework sees a returned Agent, it switches the active agent to the new one and continues the conversation with the new agent's instructions.
• Use Case: Triage agents routing users to specialized departments (e.g., "Billing", "Refunds").

3. Context Variables

Shared state passed into client.run().

• Access: Functions can declare context_variables as an argument to receive the current state.
• Updates: Functions can return a Result object to update these variables.

Key Features

Stateless Execution

The client.run() loop is stateless. It processes the current turn, handles any function calls, manages handoffs, and returns the final response along with the updated context_variables and active agent. The caller is responsible for persisting this state.

Result Object

A powerful return type for functions that allows for atomic updates:

return Result(
    value="Final answer for the user",
    agent=next_agent,           # The Handoff
    context_variables={"key": "val"} # State update
)

Comparison: Swarm vs. ADK

*Source: lit-openai-swarm*

• openai-symphony: For comparison with background, tracker-driven orchestration.