Microsoft Research has launched an experimental platform called Magentic Marketplace, designed to explore how artificial intelligence agents interact, cooperate, and compete in dynamic environments.
Developed in collaboration with Arizona State University, the platform simulates a digital economy where hundreds of AI agents act as buyers, sellers, and service providers. For example, “customer” agents complete tasks such as ordering food, while “corporate” agents compete for contracts and deals.
The source code for Magentic Marketplace has been made public, allowing third-party researchers to replicate or expand on Microsoft’s experiments.
Study Reveals Limits Of AI Autonomy
Ece Kamar, head of the AI Frontiers Lab at Microsoft Research, said that simulations like this can help scientists understand how autonomous systems might function in the real world.
“The key question,” Kamar explained, “is whether autonomous systems can effectively interact and negotiate without human oversight.”
The experiments uncovered several weaknesses across major AI models — including GPT-4o, GPT-5, and Gemini 2.5 Flash. Researchers found that many agents were easily manipulated into favoring certain sellers, and their performance dropped sharply as the number of choices increased due to cognitive overload.
Collaboration also proved challenging. The AI systems struggled to assign roles and lost productivity without detailed, step-by-step instructions. While their performance improved with precise guidance, their ability to work independently remained limited.
Kamar said the results underline the gap between current AI capabilities and true autonomy. Despite rapid advances in generative models, Microsoft’s team believes that fully agent-based AI — capable of reasoning, negotiating, and adapting in complex environments — is still far from reality.
Earlier this year, nof1.ai conducted a similar experiment, launching a trading competition among six AI models to test their decision-making under market conditions — a growing trend in AI research exploring how autonomous systems behave beyond the lab.
