Anthropic Tests AI Agent Trading: Employees Participate in Real-Money Transactions with 186 Deals Closed

AI company Anthropic recently conducted an AI agent trading experiment, building a second-hand market where AI agents act as buyers and sellers to complete transactions with real money. 69 volunteer employees participated, each receiving a $100 gift card budget to buy and sell items with colleagues.

During the experiment, 186 deals were closed with a total value exceeding $4000. Anthropic set up four independent markets: one real transaction market (supported by the company's most advanced model, with actual fulfillment), and three for academic research.

Results show that users represented by higher-level agents get better transaction outcomes, but users themselves don't notice the gap. This reveals a potential risk from AI agent capability levels—disadvantaged parties may not realize their interests are harmed. Additionally, initial instructions for agents have no obvious impact on deal probability or pricing.