AutoAgents is an experimental framework for autonomously generating and orchestrating multiple AI agents to tackle complex tasks. It is designed for researchers and developers exploring multi-agent systems and LLM-driven automation, offering a structured approach to collaborative problem-solving.

How It Works

The framework employs a hierarchical structure where a Planner agent decomposes a given goal into a series of steps. For each step, it determines necessary expert roles and defines an execution plan. Agents are then generated with specific expertise and tools. Observers monitor the process, providing reflection on agents, plans, and actions to ensure reasonableness. The system supports tool usage, primarily search tools, and allows for custom agent collections.

Quick Start & Requirements

• Installation: git clone https://github.com/LinkSoul-AI/AutoAgents followed by cd AutoAgents and python setup.py install.
• Prerequisites: OpenAI API key is mandatory. SerpAPI key is recommended for search tool functionality.
• Configuration: Set OPENAI_API_KEY and optionally OPENAI_API_BASE in config/key.yaml, config/config.yaml, or environment variables.
• Usage: Run via command line (python main.py --mode commandline ...) or as a WebSocket service (python main.py --mode service ...). Docker support is available.
• Demo: An online demo is available on HuggingFace Spaces.

Highlighted Details

• IJCAI 2024 accepted paper.
• Features autonomous generation of expert roles and execution plans.
• Includes Observer agents for process reflection and validation.
• Supports custom agent collections via AgentBank.

Maintenance & Community

The project is actively seeking collaborators for software development, documentation, and application exploration. Contact information for questions and feedback is provided via email and GitHub Issues.

Licensing & Compatibility

• License: MIT License.
• Compatibility: Permissive MIT license allows for commercial use and integration with closed-source projects.

Limitations & Caveats

Currently, the framework's tool compatibility is limited primarily to search tools. The project is described as experimental, suggesting potential for ongoing changes and undiscovered issues.