Anthropic has conducted an interesting experiment to test the feasibility of direct purchases between AI agents. In this project, called “Project Deal,” the company created a special marketplace (advertising platform) where agents act as both buyers and sellers.
According to the company, this was only a pilot test, and 69 employees voluntarily participated. Each participant was given a $ 100 budget (in the form of a gift card) to buy products from their colleagues. Nevertheless, the results were impressive: a total of 186 real deals took place, and their total value was more than $ 4,000.
Anthropic said that four different marketplace models were built as part of the experiment. One of them was a “real” model — where participants were represented by the company’s most advanced AI model, and the deals concluded were actually executed after the experiment. The other three models were for research purposes.
According to the results of the study, users represented by more advanced models achieved “objectively better results.” Interestingly, participants barely noticed this difference. This suggests that “agent quality differences” could be a problem in the future—that is, users on the weaker side may not realize that they are getting worse results.
It was also noted that the pre-instructions given to agents did not significantly affect the likelihood of a sale or price negotiations.
