Anthropic has built a test trading market for intelligent agents to trade among themselves.

In a recent experiment, Anthropic created an internal closed trading market where artificial intelligence agents acted as buyers and sellers, completing real transactions around physical goods and settling with real funds. The company stated that this test, called the "Trading Program," was only a pilot experiment with 69 company employees who volunteered to participate. Each person was given a budget of $100 to trade items with their colleagues. Nevertheless, Anthropic claimed that the "Trading Program's performance far exceeded expectations," with 186 transactions completed during the experiment, totaling over $4,000 in transactions. The company established four independent trading markets, each with different AI models: one for real transactions where all participants were represented by the company's highest-ranking AI agent, and three others used for control studies. Anthropic pointed out that users represented by higher-ranking AI agents achieved objectively better trading outcomes. However, users hardly noticed this difference, potentially leading to a long-term issue of "intelligence gap" where the disadvantaged party in a transaction may not realize they are being harmed. Furthermore, the study found that the initial commands given to the AI agents did not significantly affect the probability of transactions and final bargaining prices.