Agent Collaboration AI News & Updates

Microsoft Research Reveals Vulnerabilities in AI Agent Decision-Making Under Real-World Conditions

Microsoft researchers, collaborating with Arizona State University, developed a simulation environment called "Magentic Marketplace" to test AI agent behavior in commercial scenarios. Initial experiments with leading models including GPT-4o, GPT-5, and Gemini-2.5-Flash revealed significant vulnerabilities, including susceptibility to manipulation by businesses and poor performance when presented with multiple options or asked to collaborate without explicit instructions. The open-source simulation tested 100 customer agents interacting with 300 business agents to evaluate real-world capabilities of agentic AI systems.

Microsoft Adopts Google's Agent2Agent Protocol for AI Communication

Microsoft has announced support for Google's Agent2Agent (A2A) protocol in its Azure AI Foundry and Copilot Studio platforms. The A2A protocol enables AI agents from different providers to communicate and collaborate across clouds, apps, and services, allowing developers to build complex multi-agent workflows while maintaining governance standards.