Did OpenAI Copy Agency Swarm? In Depth Comparison

---

---

AI Summary

Video Summary: Comparison of OpenAI’s Swarm and Agency Swarm Frameworks

Fundamentals of Multi-Agent Communication

• Multi-agent communication is a primary feature with nuances in each framework.
• Agency Swarm: Agents communicate through a “send message” tool, exchanging tasks and responses.
• Swarm: Agents hand off requests without direct communication, replacing themselves in the conversation history.

Key Distinctions

• Swarm allows agents to share the same conversation thread, potentially saving on token costs and latency.
• Agency Swarm supports complex workflows with collaborative agents.
• Swarm is not based on the Assistant API, lacks built-in file search and code interpreter, and requires manual conversation history management.
• Swarm is designed to run fully on a client without sharing state between calls.

Context Variables Feature

• Swarm introduces context variables to hardcode information for reuse in function calls.
• Agency Swarm can achieve similar functionality using the “additional instructions” parameter.

Side-by-Side Comparison on a Task

• Swarm Example:
  • Customer service agent can transfer to sales or issues and repairs agents.
  • Uses a “run demo loop” in the terminal.
  • Conversation continues with the new agent in the same thread after a transfer.
• Agency Swarm Example:
  • Customer service agent sends a message to the sales agent with context.
  • Sales agent provides options and promotions, which the customer service agent relays back.
  • Agents collaborate rather than transferring the user.
  • Can activate a specific agent in the same thread using the ”@” sign.

Key Considerations for Usage

• Swarm is suitable for research and experimentation with greater control over the system.
• Agency Swarm is better for production and complex tasks requiring scalability.
• Agency Swarm allows easy addition of new agents to the system without creating multiple new tools.
• Swarm requires additional infrastructure for production, such as databases for conversation history.

Conclusion

• Swarm is recommended for learning and designing agent systems from scratch.
• Agency Swarm is recommended for building agents for real use cases and production.
• Agency Swarm has a supportive Discord community and is fully supported, unlike Swarm.

Links

• All relevant links are provided in the video description.