Microsoft has introduced a new multi-agent AI system called Magnetic-One based on a single AI model that powers multiple AI agents to complete complex tasks. The company has made the framework open-source and accessible to all developers and researchers even for commercial purposes under a custom Microsoft License. 

The blog posted by Microsoft along with the news called Magnetic-One a “high-performing generalist agentic system” that is able to use a web browser to book a ticket or buy something online, edit a document and write Python code.

The system works with a lead agent called the Orchestrator which then directs the other four agents to solve tasks. The Orchestrator plans, tracks and re-plans if errors are made while also delegating specialised agents. 

The blog noted that this is much better than single AI agent systems that are usually more rigid. On the other hand, multi-agentic frameworks can activate some agents to be added or removed even as the other agents continue with their workflows. 

Microsoft has also released a tool called AutoGenBench to evaluate how the AI agents perform. The tool comes with built-in controls for repitition and isolation to score them accurately.

AI agents are largely been predicted as the next prominent stage in AI research after AI chatbots. 

Google is reportedly working on an AI agent called Jarvis AI which will help users browse the web via Chrome.