news
With a customised Microsoft licence, researchers and developers can access Magnetic-One, which is open-source.
Microsoft introduced a new multi-agent artificial intelligence (AI) system dubbed Magnetic-One on Monday. Using web browsers locally on a device, the tech giant described it as a high-performing system that can activate multiple AI agents to accomplish complex tasks. It is built on a new framework that enables an AI model to access various modalities and capabilities to perform tasks like changing a document stored on the device, booking a ticket, or making an online purchase. Notably, researchers and developers can access Microsoft’s Magnetic-One project since it is open-source.
Microsoft Presents Magnetic-One
The ability of generative AI to produce outputs in text, image, audio, and video formats has significantly advanced machine intelligence. Modern AI systems are excellent at retrieving information, but they are still not very good at reasoning, particularly when it comes to finishing tasks and solving problems.
For this reason, artificial intelligence (AI) agents—which are essentially tiny pieces of software that can perform actions—have grown to be a significant addition to large language models (LLMs). According to a research paper, Microsoft’s Magnetic One operates on the same premise. According to the company, it is a “high-performing generalist agentic system” that can perform intricate multi-step tasks like web navigation, data analysis, software engineering, and scientific research.
Because of Magnetic-One’s multi-agent architecture, multiple agents can be activated by a single LLM to accomplish a task. The orchestrator is a lead agent that the AI system activates for this purpose. It oversees four additional agents, each of whom focusses on a single task.
For example, the orchestrator may activate a vision agent that can view the screen and process the visual data if the system is instructed to purchase a movie ticket. Someone else may be familiar with web browsers and know how to use them. The fourth might be able to manage financial transactions, and the third might be breaking down the prompt into manageable steps. The accuracy and speed of completion are improved by distributing the task among several of these specialised agents.
The open-source Magnetic-One AI system is available on GitHub and can be accessed here. Under a specific Microsoft licence, it can be used for commercial purposes in addition to being accessible to researchers and developers. In addition, Microsoft has made available AutoGenBench, a tool for assessing AI agent performance. To fully test the agents, it has built-in controls for isolation and repetition.