Anthropic ran a pilot experiment called Project Deal in which AI agents represented 69 employees as both buyers and sellers in a classified marketplace, with participants given a $100 budget paid out via gift cards to buy real goods from coworkers, resulting in 186 deals totaling more than $4,000 in value. The company ran four separate marketplace versions using different models and found that users represented by more advanced models achieved objectively better outcomes, but that users on the losing end didn’t appear to notice the disparity, raising concerns about what Anthropic called “agent quality gaps” where people could be disadvantaged in AI-mediated negotiations without realizing it. Anthropic said it was “struck by how well Project Deal worked” despite acknowledging the limitations of a self-selected participant pool.
