AWorld is a multi-agent runtime designed for building, evaluating, and training collaborative AI agents that can self-improve. It targets researchers and developers working on complex multi-agent systems (MAS), foundation model training, and product prototype verification, offering a flexible framework for agent orchestration and environment interaction.

How It Works

AWorld facilitates agent collaboration through a structured runtime that handles agent construction, topology orchestration, and environment management. It supports integrating various model services and custom tools via protocol encapsulation, enabling agents to communicate and interact within defined environments. This design allows for the creation of diverse runtime configurations to discover and push the limitations of AI models.

Quick Start & Requirements

• Installation: git clone https://github.com/inclusionAI/AWorld, cd AWorld, python setup.py install
• Prerequisites: Python >= 3.11. OpenAI API key and potentially other API keys for specific MCP tools (e.g., amap-amap-sse). Docker image available for quick GAIA evaluation.
• Demo: Examples for GAIA agent runtime and multi-agent collaboration are provided.

Highlighted Details

• Ranks 3rd on the GAIA benchmark (69.7 avg, Pass@1 = 58.8), leading among open-source frameworks.
• Supports agent self-improvement through diverse runtime environments.
• Offers integration with RL learning frameworks like ms-swift for scalable agent training.
• Provides demos for single-agent (browser/phone) and multi-agent (cooperative/competitive) scenarios.

Maintenance & Community

• Active development with recent updates in May-July 2025.
• Community channels include Discord and Twitter.
• Contact email provided for academic citations and inquiries.

Licensing & Compatibility

• Licensed under the MIT License, permitting commercial use and integration with closed-source projects.

Limitations & Caveats

• Deploying online distributed environments for high-concurrency rollouts requires contacting a specific email address for assistance.
• Some features, like Function call runtime and Deep Search runtime, are noted as "to be released."