Downcodes editor reports: The Microsoft research team recently released a new multi-agent system called Magnetic-One, which is an open source project designed to improve the work efficiency of individuals and enterprises. The system allows one AI model to drive multiple assistant agents to collaboratively complete complex multi-step tasks. Its core is a commander agent, which is responsible for managing and coordinating four other different types of auxiliary agents: Websurfer, FileSurfer, Coder and ComputerTerminal. These agents are respectively responsible for web browsing, file processing, code writing and terminal control, and jointly complete complex tasks.
Recently, the Microsoft research team released a new multi-agent system called Magnetic-One. The framework allows an AI model to drive multiple assistant agents to collaboratively complete complex multi-step tasks, aiming to improve the work efficiency of individuals and enterprises.
Microsoft calls Magnetic-One a universal intelligent system that can realize its long-term vision of enhancing productivity and changing lifestyles.
Magnetic-One is an open source project that can be developed for commercial use by researchers and developers under a custom Microsoft license. At the same time, Microsoft also released an open source agent evaluation tool called AutoGenBench, designed to test the effect of the agent system. This tool is based on the previously released Autogen framework and supports communication and cooperation between multiple agents.
The core of Magnetic-One lies in an Orchestrator, which is responsible for managing and coordinating four different auxiliary agents. In addition to the commander, the system also includes the following four types of agents:
1. **Websurfer Agent**: Able to control Chromium-based browsers, perform web search, click and input, and even summarize web content.
2. **FileSurfer Agent**: used to read local files, list directories and browse folders.
3. **Coder agent**: Responsible for writing code, analyzing information from other agents and creating new projects.
4. **ComputerTerminal agent**: Provides a console for the program execution of the Coder agent.
The commander agent not only formulates a task execution plan, but also tracks the progress of each agent in real time and reassigns it when errors occur. If the agent encounters obstacles on a certain task, the commander can adjust the plan to ensure the successful completion of the task.
Microsoft used OpenAI's GPT-4o when developing Magnetic-One. However, the system is independent of large language models, and the researchers recommend using a powerful inference model as the commander agent. Magnetic-One supports a variety of model combinations, and developers can flexibly choose according to their needs.
As enterprises become increasingly interested in multi-agent deployment, Microsoft's Magnetic-One system may become an important part of the evolution of the agent framework. Other companies such as OpenAI and CrewAI are also launching similar agent management frameworks, and companies are still exploring when choosing the best multi-agent solutions.
Official introduction: https://www.microsoft.com/en-us/research/articles/magentic-one-a-generalist-multi-agent-system-for-solving-complex-tasks/
The emergence of Magnetic-One has brought new possibilities to the development of multi-agent systems, and its open source features also facilitate developers' research and application. In the future, as technology continues to mature, it is believed that multi-agent systems will play an important role in more fields. Looking forward to the emergence of more innovative applications!