Microsoft Introduces Magnetic-One Generalist Multi-Agent AI System That Can Complete Complex Tasks

Microsoft introduced a new multi-agent artificial intelligence (AI) system dubbed Magnetic-One on Monday. The tech giant called it a high-performing system that can activate multiple AI agents to complete complex tasks via web browsers on locally on a device. It is based on a new framework that allows an AI model to access multiple modalities and capabilities to complete tasks such as booking a ticket, purchasing a product online, or editing a document stored on the device. Notably, Microsoft's Magnetic-One is an open-source project and is accessible to researchers and developers.

Microsoft Introduces Magnetic-One

Generative AI has taken a huge leap in machine intelligence and its capability to generate outputs across text, images, audio, and video formats. However, while modern AI systems are great at retrieving information, they still remain poor at reasoning, especially when it comes to solving problems and completing tasks.

This is why AI agents, which can be understood as miniature software capable of executing an action, have become an important extension of large language models (LLMs). Microsoft's Magnetic-One also works on the same principle, as detailed in a research paper. The company describes it as a “high-performing generalist agentic system” designed to complete complex multi-step tasks such as software engineering, data analysis, scientific research, and web navigation.

Magnetic-One has a multi-agent architecture, which means one LLM can activate several agents to complete a task. For this, the AI system activates a lead agent dubbed the Orchestrator. It directs four other agents where each agent specialises in one task.

The workflow of Magnetic One
Photo Credit: Microsoft

For instance, if the system is asked to book a ticket for a movie, the Orchestrator could trigger a vision agent that can look at the screen and process the visual information. Another might have knowledge of web browsers and can handle its navigation. The third could be breaking down the prompt into actionable steps, and the fourth might be able to handle financial transactions. By dividing the task among multiple such specialised agents, both the accuracy and speed of completion is increased.

The open-source Magnetic-One AI system is available on GitHub and can be accessed here. It is available to researchers and developers, and can also be used for commercial purposes under a custom Microsoft licence. Alongside, Microsoft has also released AutoGenBench, which is a tool that evaluates the performance of AI agents. It comes with built-in controls for repetition and isolation to thoroughly test the agents.

(Except for the headline, this story has not been edited by NDTV staff and is published from a press release)