In a recent experiment, Anthropic created a confidential marketplace where AI agents represent both buyers and sellers, conducting real transactions for real goods and real money.
The company confirmed that the test, called Project Deal, was only a “pilot experiment with a self-selected participant pool” of 69 Anthropic employees who were given a budget of $100 (paid for in gift cards) to buy things from co-workers.
Nevertheless, Anthropic said it was “shocked by how well the project deals worked,” with 186 transactions totaling more than $4,000.
The company said it actually operated four different marketplaces with different models. One was “real” (all were represented by the company’s cutting-edge models and were actually approved for trading after testing), and the other three were for research purposes.
Apparently, users get “objectively better results” when they are represented by a more sophisticated model, Anthropic said. However, users do not seem to be aware of this disparity, increasing the possibility of a “difference in the quality of agents” in which “people on the losing side may not be aware that their situation is getting worse.”
It also appears that the initial instructions given to the agents did not affect the likelihood of a sale or the negotiated price.
