Recently, Microsoft Research released a paper discussing an innovative multi-agent collaboration framework: Magentic-One.
It is a general-purpose agent system designed to efficiently handle complex workflows by coordinating multiple dedicated AI agents.
The system architecture of Magnetic-One is as follows, where the red part represents the brain of the entire framework, the Orchestrator, responsible for task management and planning.
-
Websurfer Agent: Performs web searches, clicks, inputs, and even summarizes web content. -
FileSurfer Agent: Used for reading local files, listing directories, and browsing folders. -
Coder Agent: Responsible for writing code, analyzing information from other agents, and creating new projects. -
ComputerTerminal Agent: Provides a console for the execution of programs by the Coder agent.
The brain not only formulates task execution plans but also tracks the progress of each agent in real-time and reassigns tasks in case of errors.
It implements two loops, where the external loop is responsible for task management and planning, while the internal loop mainly ensures that tasks can be completed smoothly. Both loops share information to achieve flexible and reliable multi-agent collaboration. Additionally, Magentic-One achieves automatic error correction by obtaining error code distribution and error code heat maps.
Magnetic-One is a powerful and flexible multi-agent collaborative framework, expected to play an important role in the future of intelligent fields.
Paper link:
https://arxiv.org/abs/2411.04468