MetaAgent: Automatically Building Multi-Agent System based on Finite State Machine

🌐 [Website]](https://mercury7353.github.io/MetaAgent-Static-Demo/) 🔗 Paper

🎆 News

• [2025/5] Accepted to ICML2025.

👀 Overview

MetaAgent is a framework that can be used to build your own multi-agent system automatically by one line of prompt.
The Open-Sourced framework

The Multi-Agent System (MAS) is based on the Finite State Machine (FSM), which is shown in the figure.

Auto-generation of Multi-Agent System

• Given a general task description(eg. Build a multi-agent system for software development), MetaAgent can automatically generate a multi-agent system with several agents.
• Unlike other auto-generation framework, MetaAgent can generate a multi-agent system without external training data.
• The generated multi-agent system is also able to solve every cases of the given task domain.

Why Finite State Machine?

The finite state machine has several features:

🔧 Tool Enabled: The agent can use tools to help it complete tasks.
🔄 State Traceback: The agent can traceback the state of the system when the task fails.

📹 Demo Video

1017.mp4

🚀 Quick Start

Install

To set up the environment of MetaAgent, please run the following command:

pip install -r requirements.txt

Setup your API-Keys

export OPENAI_API_KEY='your-openai-api-key'
export SERP_API_KEY=''
... (Other API Keys for tools)

Run

• In the code base:

cd autodesign
python building_pipeline.py

cd evaluate
python general_test.py # test any input on the designed FSM.

• In the interface:

cd web
python app.py

Customize Your own FSM

• Use baseclass/FSM_GEN.py to auto-design your FSM
• Use autodesign/Optimization.py to optimize the FSM
• Use autodesign/evaluate/general_test.py to test the FSM

🤝 Support

Contact Us

📮 Email: zhangyaolun5@gmail.com