On Monday, Microsoft unveiled a new multi-agent artificial intelligence (AI) system named Magnetic-One. The company described this advanced system as capable of activating several AI agents to fulfill complex tasks using web browsers or directly on a device. This innovation leverages a new framework that enhances the AI model’s abilities across multiple modalities, enabling it to perform tasks such as booking tickets, purchasing items online, or editing documents stored locally. Importantly, Magnetic-One is an open-source initiative, making it available to researchers and developers.
Microsoft Unveils Magnetic-One
The field of generative AI has dramatically progressed in its ability to produce outputs in various formats, including text, images, audio, and video. Despite these advancements, contemporary AI systems still struggle with reasoning capabilities, particularly in problem-solving and task completion.
Consequently, AI agents—essentially compact software solutions designed to perform specific actions—have emerged as a crucial extension of large language models (LLMs). Magnetic-One operates under the same principle, according to a research paper released by Microsoft. The organization characterizes it as a “high-performing generalist agentic system,” developed to undertake intricate multi-step tasks such as software engineering, data analysis, scientific research, and web navigation.
Magnetic-One utilizes a multi-agent architecture, allowing one LLM to mobilize several agents to tackle a task. The system identifies a lead agent called the Orchestrator, which coordinates the activities of four additional agents, each specialized in a specific area.
The workflow of Magnetic One
Photo Credit: Microsoft
For example, if tasked with booking a movie ticket, the Orchestrator could activate a vision agent to analyze visual information on the screen. Another agent might specialize in web navigation, while a third breaks the request into actionable steps, and a fourth handles the financial transaction. By distributing the workload among multiple specialized agents, the system enhances both the accuracy and speed of task completion.
The open-source Magnetic-One AI system is hosted on GitHub and can be accessed here. It is available to researchers and developers for commercial use under a custom Microsoft license. Additionally, Microsoft has launched AutoGenBench, a tool designed to assess AI agent performance, equipped with built-in controls for repetition and isolation to thoroughly evaluate the agents.
(Except for the headline, this story has not been edited by NDTV staff and is published from a press release)
