Microsoft recently filed a patent application with the US Patent and Trademark Office for a system designed to coordinate competing AI agents. This system quantitatively measures the value of agent-generated responses to improve operational performance. The initiative supports Microsoft’s goal of building frameworks that enable specialized agents to collaborate on complex tasks, rather than relying solely on a single agent.
Main Points From the Patent and Microsoft’s AI Agent Strategy
- Conflict and evaluation: The patent outlines methods to measure the value of responses from multiple agents operating simultaneously, even when those responses conflict.
- Orchestration: Microsoft uses an orchestrator agent to assign, monitor, and adjust tasks for subordinate agents, ensuring coordination and goal consistency throughout the multi-agent system.
- Multi-agent systems (magnetic one): the magnetic-one framework manages specialized agents that perform tasks such as web browsing, file management, and coding, as announced in November 2024.
- Efficiency measures: Microsoft is creating quantitative tools to precisely assess agent efficiency and accuracy. The company notes that offering too many response options can reduce agent performance, a problem recognized in research as overwhelming agent attention.
- Agent boss mindset: Microsoft envisions a future where employees manage teams of AI agents. For instance, a sales engine could prompt an inventory agent to automate a workflow.
- Safety and security. Microsoft is building lists of potential failure modes to keep these systems secure and to manage agents that exhibit unexpected behavior.
These patent activities and related research drive Microsoft’s strategy to create autonomous agents capable of machine reasoning, environmental navigation, logical problem-solving, and operating independently in intricate business and personal situations.
Microsoft has launched a new multi-agent AI system called Magnetic One. It uses a single AI model to run multiple agents that can handle complex tasks. The company has made the framework open source so that any developer or researcher can use it, including for commercial projects, under a Microsoft-custom license.
Microsoft calls Magnetic One a high-performance, generalist agentic system that operates a web browser, reserves tickets or makes purchases, modifies documents, and generates Python code.
The system features a lead agent, the orchestrator, which assigns tasks to four other agents for completion. The orchestrator handles planning, oversight, error correction, and task delegation to support agents.
According to the blog, this architecture outperforms inflexible single-agent systems. Multi-agent frameworks allow agents to be added or removed without disrupting operations.
Microsoft also released a tool called AutoGenbench to measure AI agent performance, including controls for redundancy and separation to guarantee reliable evaluation.
Experts expect AI agents to drive the next major advance in AI research after chatbots.
Reports say Google is developing Jarvis AI to help users browse the Chrome web browser.
Source: Microsoft unveils open-source multi-agent AI system Magnetic-One
