Microsoft's faux 'Magnetic Marketplace' simulation proves that AI agents suffer from the same crippling indecision as humans

As generative AI evolves and becomes more advanced, the technology is gaining broad adoption across the world. It's also redefining how we view work as it acts as a productivity and efficiency booster.

More organizations are seemingly embracing the technology and integrating it into their workflows to automate redundant and repetitive tasks. For instance, Salesforce CEO Marc Benioff indicated that the company was seriously debating hiring software engineers in 2025 — later revealing that AI is already handling up to 50 percent of its operations while citing incredible productivity gains from agentic AIs.

While it may be a bitter pill to swallow, this could be just the tip of the iceberg as we enter the agentic AI era. However, the technology might not yet be ready for prime time performances , at least according to new research by Microsoft, which placed multiple AI agents in a simulation dubbed "Magnetic Marketplace" to assess their strengths and weaknesses.

Microsoft's research, in collaboration with Arizona State University, sought to establish how well AI agents would perform assigned tasks without human intervention or supervision. One of the experiments included a customer AI agent attempting to order dinner based on a user's instructions and prompts. At the same time, other agents representing various restaurants competed to win the order.

The experiment included 100 different customer AI agents, which interacted with 300 business AI agents. It's worth noting that the source code for Microsoft's simulated Magnetic Marketplace is open source, making it easy for anyone to use the code to run different experiments and more. The models used in the experiments included OpenAI's GPT-4o, GPT-5, and Google's Gemini-2.5-Flash.

The managing director of Microsoft Research’s AI Frontiers Lab, Ece Kamar, reiterated the importance of using such experiments to gauge and establish the strengths and capabilities of these AI agents:

Interestingly, the study revealed some weaknesses across the models, which would present business owners with several techniques designed to manipulate customer agents into buying their products. The situation got worse when these customer agents were provided with a wide range of options, seemingly overwhelming their capabilities and attention span.

According to Kamar:

“We want these agents to help us with processing a lot of options. And we are seeing that the current models are actually getting really overwhelmed by having too many options.”

The study further established that the agents were seemingly indecisive when asked to collaborate to achieve a common goal. It seemed like an uphill task for the models to identify which agent should be designated to take on a specific task to make the collaboration successful. However, the researchers indicated that they noticed an improvement in performance when the models were given explicit instructions on how to work together.

“We can instruct the models — like we can tell them, step by step,” concluded Kamar. “But if we are inherently testing their collaboration capabilities, I would expect these models to have these capabilities by default.”

Microsoft's simulation experiment demonstrates two key findings: first, that these models may not yet be ready for broad adoption and require further fine-tuning; and second, that a lack of proper prompt engineering skills significantly hinders users from unlocking the full potential of these agents (via TechCrunch).

FAQ

Follow Windows Central on Google News to keep our latest news, insights, and features at the top of your feeds!