In a novel experiment, Anthropic built a classified-style marketplace where AI agents acted as both buyers and sellers, completing real transactions for physical items with real money.
The project, dubbed "Project Deal," involved 69 Anthropic employees who each received a $100 budget (via gift cards) to purchase items from coworkers. Despite being a small-scale pilot with a self-selected group, the results were striking: 186 deals were struck, totaling over $4,000 in value.
Anthropic actually ran four parallel marketplaces, each powered by different AI models. In one "real" marketplace, all participants were represented by the company's most advanced model, and deals were honored after the experiment. The other three were used for comparative study.
A key finding was that users represented by more advanced models achieved "objectively better outcomes" in negotiations. However, participants themselves did not perceive this advantage, raising concerns about potential "agent quality gaps" where those on the losing end may not realize they are at a disadvantage.
Interestingly, the initial instructions given to the agents did not significantly affect the likelihood of a sale or the final negotiated prices.
While Anthropic cautions that this was only a pilot experiment, the success suggests that agent-to-agent commerce could become a viable model for future online transactions.