Microsoft Launches Magnetic-One: AI System to Transform Business Efficiency

Microsoft’s new open-source AI system, Magnetic-One, lets AI agents automate difficult, multi-step tasks on the web and in files, which makes businesses more efficient.

Microsoft researchers have developed a new open-source multi-agent AI system to help businesses automate challenging tasks that typically require human assistance.

The Magnetic-One project aims to manage a wide range of complex and limitless tasks in both web and file-based environments.

Microsoft said the new system will let AI agents talk to people and perform complicated tasks that require multiple steps. The company called this a big step forward for AI development.

“It’s the difference between generative AI recommending dinner options to agentic assistants that can autonomously place your order and arrange delivery,” said Microsoft researchers. “It’s the shift from summarizing research papers to actively searching for and organizing relevant studies in a comprehensive literature review.”

Magnetic-One uses a multi-agent system, similar to Salesforce’s Agentforce, to automate and carry out tasks. The “Orchestrator” agent, located in the middle of the system, controls and coordinates the four other agents, namely WebSurfer, FileSurfer, Coder, and ComputerTerminal.

Each agent is in charge of a specific task, such as browsing the web, working with files, writing code, or using the command line. The orchestrator assigns subtasks on the fly, keeps an eye on progress, and changes its plan as needed to finish complicated tasks with little help from people.

“The orchestrator plans tracks progress, and reorganizes to recover from errors while directing specialized agents to perform tasks like operating a web browser, navigating local files, or writing and executing Python code,” researchers with Microsoft AI said.

Microsoft Launches Magnetic-One as Open Source

Microsoft is making Magnetic-One available as an open-source project for developers and researchers. Although the system is excellent at many things, it is still not as effective as a human being and may still make mistakes.

Microsoft said that as agentic systems get smarter, risks like doing things you didn’t mean to or abusing them could grow.

Microsoft knows that agentic AI is still in its early stages and is counting on the public to help solve these problems and make sure that future systems are both useful and safe.

To help with this, the company is also releasing AutoGenBench, an evaluation tool made to test agentic tasks thoroughly. It has built-in controls to reduce unwanted side effects by isolating and repeating tasks over and over again.

“AutoGenBench makes agent-based evaluation easier and lets you add new benchmarks.” “AutoGenBench lets us test Magnetic-One on several different benchmarks,” Microsoft wrote.

When selecting benchmarks, they must be challenging, require numerous steps, and at least partially involve planning and the use of tools such as web browsers to interact with real or fake web pages. This paper looks at three benchmarks GAIA, Assistant Bench, and Web Arena meet this requirement.