Microsoft’s new Magnetic-One system directs multiple AI agents to complete user tasks

Enterprises looking to deploy multiple AI agents often need to implement a framework to manage them. 

To this end, Microsoft researchers recently unveiled a new multi-agent infrastructure called Magnetic-One that allows a single AI model to power various helper agents that work together to complete complex, multi-step tasks in different scenarios. Microsoft calls Magnetic-One a generalist agentic system that can “fully realize the long-held vision of agentic systems that can enhance our productivity and transform our lives.”

The framework is open-source and available to researchers and developers, including for commercial purposes, under a custom Microsoft License. In conjunction with the release of Magnetic-One, Microsoft also released an open-source agent evaluation tool called AutoGenBench to test agentic systems, built atop its previously released Autogen framework for multi-agent communication and cooperation.

The idea behind generalist agentic systems is to figure out how autonomous agents can solve tasks that require several steps to finish that are often found in the day to day running of an organization or even an individual’s daily life. 

From the examples Microsoft provided, it looks like the company hopes Magnetic-One fulfills almost mundane tasks. Researchers pointed Magnetic-One to tasks like describing trends in the S&P 500, finding and exporting missing citations, and even ordering a shawarma. 

How Magnetic-One works

Magnetic-One relies on an Orchestrator agent that directs four other agents. The Orchestrator not only manages the agents, directing them to do specific tasks, but also redirects them if there are errors.

The framework is composed of four types of agents other than the Orchestrator:

• Websurfer agents can command Chromium-based web browsers and navigate to websites or perform web searches. It can also click and type, similar to Anthropic’s recently released Computer Use, and summarize content. 
• FIleSurfer agents read local files list directories and go through folders.
• Coder agents write codes, analyze information from other agents and create new artifacts.
• ComputerTerminal provides a console where the Coder agent’s programs can be executed. 

The Orchestrator directs these agents and tracks their progress. It starts by planning how to tackle the task. It creates what Microsoft researchers call a task ledger that tracks the workflow. As the task continues, the Orchestrator builds a progress ledger “where it self-reflects on task progress and checks whether the task is completed.” The Orchestrator can assign an agent to complete each task or update the task ledger. The Orchestrator can create a new plan if the agents remain stuck. 

“Together, Magentic-One’s agents provide the Orchestrator with the tools and capabilities that it needs to solve a broad variety of open-ended problems, as well as the ability to autonomously adapt to, and act in, dynamic and ever-changing web and file-system environments,” the researchers wrote in the paper. 

While Microsoft developed Magnetic-One using OpenAI’s GPT-4o — OpenAI is after, all a Microsoft investment — it is LLM-agnostic, though the researchers “recommend a strong reasoning model for the Orchestrator agent such as GPT-4o.” 

Magnetic-One supports multiple models behind the agents, for example, developers can deploy a reasoning LLM for the Orchestrator agent and a mix of other LLMs or small language models to the different agents. Microsoft’s researchers experimented with a different Magnetic-One configuration “using OpenAI 01-preview for the outer loop of the Orchestrator and for the Coder, while other agents continue to use GPT-4o.”

The next step in agentic frameworks

Agentic systems are becoming more popular as more options to deploy agents, from off-the-shelf libraries of agents to customizable organization-specific agents, have arisen. Microsoft announced its own set of AI agents for the Dynamics 365 platform in October. 

Tech companies are now beginning to compete on AI orchestration frameworks, particularly systems that manage agentic workflows. OpenAI released its Swarm framework, which gives developers a simple yet flexible way to allow agents to guide agentic collaboration. CrewAI’s multi-agent builder also offers a way to manage agents. Meanwhile, most enterprises have relied on LangChain to help build agentic frameworks. 

However, AI agent deployment in the enterprise is still in its early stages, so figuring out the best multi-agent framework will continue to be an ongoing experiment. Most AI agents still play in their playground instead of talking to agents from other systems. As more enterprises begin using AI agents, managing that sprawl and ensuring AI agents seamlessly hand off work to each other to complete tasks is more crucial.