Microsoft unveiled a cutting-edge multi-agent artificial intelligence (AI) system named Magnetic-One on Monday. The tech giant has referred to it as a high-performing system capable of activating multiple AI agents simultaneously to carry out intricate tasks either through web browsers or directly on a device. Smoothly put, this innovation utilizes a novel framework that empowers an AI model to tap into various modalities and functionalities in order to carry out tasks like booking a ticket, buying a product online, or editing a document saved on the device.
Notably, Microsoft’s Magnetic-One is an open-source project and is accessible to researchers and developers. Microsoft has unveiled the groundbreaking Magnetic-One. Generative AI has made significant strides in machine intelligence, enhancing its ability to produce varied outputs spanning text, images, audio, and video formats. Modern AI systems excel in retrieving information, yet they struggle with reasoning, particularly in problem-solving and task completion.
This is the reason AI agents – seen as miniature software able to perform actions – have become a vital extension of large language models (LLMs). Microsoft’s Magnetic-One operates on the same principle, as outlined in a research paper. The company portrays it as a “high-performing generalist agentic system” crafted to tackle intricate multi-step tasks like software engineering, data analysis, scientific research, and web navigation.
Magnetic-One employs a multi-agent architecture, allowing a single LLM to engage multiple agents in order to accomplish a task. For this purpose, the AI system triggers a key agent known as the Orchestrator. It guides a team of four agents, each specializing in a specific task. For example, suppose the system needs to reserve a movie ticket. The Orchestrator may activate a vision agent to analyze the screen and interpret the visual data.
Another individual may possess expertise in web browsers and is capable of navigating through them.
The third task involves breaking down the prompt into actionable steps, while the fourth task focuses on handling financial transactions. By assigning the task to several specialized agents, both accuracy and speed of completion are enhanced. The open-source Magnetic-One AI system is conveniently available on GitHub, accessible through this link. It is accessible to researchers and developers. Additionally, it can be utilized for commercial purposes with a customized Microsoft license.
Microsoft has recently launched AutoGenBench, a tool designed to assess the performance of AI agents. It is equipped with embedded controls for repetition and isolation to ensure comprehensive testing of the agents.
