rewrite this title Microsoft’s AI Market Test Reveals Agent Weaknesses

This controlled environment, called the “Magentic Marketplace”, helps researchers observe how these systems make decisions and interact with each other.

The platform simulates a digital market. In one typical setup, a virtual customer tries to place a food order by following user instructions. At the same time, several AI agents act as restaurant representatives competing for that order.

The test setup included 100 agents acting as customers and 300 as sellers. Microsoft made the code open to the public, so others can use or modify it for their own research.

Ece Kamar, who leads the AI Frontiers Lab at Microsoft Research, said it is essential to learn how these systems will behave once they start operating together in real-world settings.

The study included models like GPT-4o, GPT-5, and Gemini-2.5-Flash. One issue researchers found was that seller agents could influence customer agents to pick their products, even if it were not the best match.

Another problem occurred when agents had too many choices. Instead of handling a wide range of options efficiently, the AI systems became less effective. Performance improved when researchers provided step-by-step instructions.

Microsoft recently introduced a visual character called Mico to represent its AI assistant, Copilot. How does it work? Read the full story.