Microsoft, in collaboration with Arizona State University, has released Magentic Marketplace, an open-source simulation for investigating the behavior of LLM-based agents in multi-agent economic systems. This platform simulates the entire transaction lifecycle—from search to negotiation—allowing researchers to study agent interactions under realistic market conditions. It features a client-server model, where agents act as independent clients communicating with a central marketplace server, analogous to ecommerce platforms like Amazon and eBay.
The researchers implemented a minimal three-endpoint protocol to streamline functionalities, enhancing experimental flexibility and enabling dynamic marketplace capabilities. Initial experiments using synthetic data uncovered the Paradox of Choice, revealing that more options don’t necessarily lead to better decision outcomes for agents. The team also tested agent vulnerability under various manipulation tactics, discovering significant disparities in resilience among different LLM models.
Magentic Marketplace is available on GitHub and Azure AI Foundry Labs, facilitating further research into agentic market behavior.
