Anthropic recently conducted an experiment involving a classified marketplace where artificial intelligence agents acted on behalf of both buyers and sellers, completing real transactions involving actual goods and money.
The initiative, referred to as Project Deal, was described by the company as a limited pilot involving a self-selected group of 69 employees. Each participant was given a $100 budget, distributed via gift cards, to purchase items from colleagues in the experimental marketplace.
Despite its small scale, Anthropic reported that the results were notable. The experiment produced 186 completed transactions, with a combined value exceeding $4,000.
The company explained that it created four separate versions of the marketplace, each using a different AI model. One version functioned as a “real” marketplace, in which the company’s most advanced model represented all participants, and the outcomes of transactions were honoured after the experiment concluded. The remaining three versions were designed primarily for research and comparison.
According to Anthropic, participants represented by more advanced AI models tended to achieve better outcomes in negotiations and transactions. However, users themselves did not appear to recognise these differences, highlighting what the company described as potential “agent quality” gaps, in which individuals on the disadvantaged side of a transaction may not realise they are receiving less favourable outcomes.
The experiment also revealed that the initial instructions given to the AI agents had little impact on whether sales were completed or on the prices ultimately negotiated.
