To efficiently manage multiple AI agents deployed within enterprises, Microsoft researchers have recently launched Magnetic-One, a multi-agent infrastructure. This framework enables a single AI model to oversee numerous auxiliary agents that collaborate to execute complex, multi-step tasks across various scenarios.

Magnetic-One is an open-source universal agent system released under a customized Microsoft license, making it available for researchers and developers, including for commercial use. Additionally, Microsoft has introduced AutoGenBench, an open-source agent evaluation tool built on its previously released Autogen framework, designed to test the communication and collaboration capabilities of multiple agents.

The core objective of the universal agent system is to address tasks that require multiple steps to complete, which are common in organizational operations and everyday personal activities. According to Microsoft's examples, Magnetic-One is intended to handle routine tasks such as describing S&P 500 trends, locating and exporting missing citations, and even ordering shawarma.

The Magnetic-One framework includes an agent called Orchestrator, which directs four other types of agents: the Websurfer agent controls a Chromium-based web browser for navigation and searches, including clicking, inputting, and summarizing content; the FileSurfer agent manages reading local files, listing directories, and browsing folders; the Coder agent is responsible for writing code, analyzing information provided by other agents, and creating new files; the ComputerTerminal provides a console for executing programs developed by the Coder agent.

The Orchestrator agent is responsible for planning tasks, creating workflow logs, and tracking progress during task execution. If an agent encounters difficulties, the Orchestrator will reassign the task or develop a new plan.

Although Magnetic-One is developed based on OpenAI's GPT-4, it is model-agnostic. Developers can deploy a powerful reasoning language model like GPT-4 for the Orchestrator agent, while other agents can utilize different or smaller language models.

As the deployment of AI agents within enterprises continues to grow, managing these agents and ensuring their seamless collaboration to complete tasks becomes increasingly important. Currently, several technology companies are competing in the AI orchestration framework market, including OpenAI's Swarm framework, CrewAI's multi-agent builder, and the widely used LangChain. However, the deployment of AI agents in businesses is still in its early stages, and the optimal multi-agent framework is yet to be determined.