VRSEN/agency-swarm

What it solves

Agency Swarm simplifies the creation of multi-agent AI applications by organizing them like real-world organizational structures. It provides a structured way to define agent roles, manage how they communicate, and integrate tools, reducing the complexity of orchestrating multiple AI agents working together on complex tasks.

How it works

Built on top of the OpenAI Agents SDK, the framework allows users to define individual Agent objects with specific roles, instructions, and tools. These agents are then grouped into an Agency, where the developer defines directional communication_flows (using a > operator) to specify which agents can initiate conversations with others. It supports type-safe tool creation via Pydantic and can convert OpenAPI schemas directly into agent tools. State persistence is handled through customizable callbacks for saving and loading conversation threads.

Who it’s for

Developers building production-ready multi-agent systems who want an intuitive, organizational-style approach to agent orchestration and a flexible way to integrate various LLM backends (via LiteLLM).

Highlights

• Organizational Orchestration: Uses a directional communication flow system to model agent interactions after a company hierarchy.
• Type-Safe Tooling: Implements Pydantic models for automatic argument validation of agent tools.
• OpenAPI Integration: Ability to convert OpenAPI schemas directly into usable tools for agents.
• Flexible Model Support: Compatible with OpenAI native models and other providers like Anthropic, Google, and Grok via LiteLLM.
• Built-in Demos: Includes a Web UI copilot demo and a Terminal UI (TUI) for rapid testing and interaction.