In a recent experiment, Anthropic created a classified marketplace where AI agents represent both buyers and sellers, conducting real transactions for real goods and real money.
The company acknowledged that the test, called Project Deal, was only a “pilot experiment using a hand-selected participant pool” of 69 Anthropic employees who were given a budget of $100 (paid via gift card) to buy items from co-workers.
Nonetheless, Anthropic said it was “surprised by how well Project Deal has worked,” with 186 deals closed, with a total value of more than $4,000.
The company said it actually operated four separate marketplaces with different models. One is “real” (where everyone is represented by the company’s most advanced model and deals are actually made after experiments) and the other three are for research.
Clearly, when users are represented by a more advanced model, they get “objectively better results,” Anthropic said. But users didn’t seem to notice this difference, raising the possibility of an “agent quality’ gap” where “people on the victim’s side may not realize they are in a worse position.”
Additionally, the initial instructions given to the agent do not appear to have any effect on the likelihood of a sale or negotiated price.
