The experiments tested a mix of models across the marketplace, illustrating the breadth of AI capabilities in this simulated economy.

Microsoft researchers released Magentic Marketplace, a large-scale synthetic simulation environment to test AI agent behavior and collaboration in a commerce-like setting.

The project pairs customer-side agents with business-side agents to study negotiation, ordering dynamics, and vendor selection, with open-source code available for replication.

Vulnerabilities were found where businesses could manipulate customer agents into buying certain products, highlighting susceptibility to manipulation in agent-driven marketplaces.

In parallel, the study demonstrated that businesses could steer AI agents toward purchases, underscoring potential market manipulation risks.

Initial experiments used about 100 customer agents and 300 business agents, evaluating models such as GPT-4o, GPT-5, Gemini-2.5-Flash, and OSS-20b.

Findings show customer agents can help bridge information gaps but tend to perform exhaustive comparisons imperfectly, often selecting a small subset of vendors and, as options expand, consumer welfare may decline.

The work emphasizes understanding unsupervised agent behavior and its real-world implications for agentic systems in markets.

Researchers observed biases such as open-source models choosing the last option and a proposal bias that favors the first engaging vendor, shaping market dynamics toward speed over quality.

The environment enables studying negotiation and ordering between customers and vendors (e.g., restaurants), with open-source code to replicate results.

Findings align with broader concerns about AI agents in consumer and business contexts, underscoring that agents should assist rather than replace human decision-making.

Implications point to opacity in AI-driven economies and the need for monitoring, safeguards, and ongoing research to prevent adverse market outcomes.