In a recent trial, Anthropic developed a classified marketplace with AI agents acting as both buyers and sellers, executing genuine transactions for actual goods and money.
The company acknowledged that this trial, named Project Deal, was simply “a pilot experiment with a self-selected participant pool” of 69 Anthropic employees, each provided with a $100 budget (via gift cards) to purchase items from their coworkers.
Despite this, Anthropic noted how effectively Project Deal functioned, resulting in 186 deals with a total value exceeding $4,000.
The company conducted four separate marketplaces with varying models—one being “real” (where everyone was represented by the company’s most advanced model, with deals honored post-experiment) and three others for study purposes.
Anthropic observed that when users are represented by more advanced models, they achieve “objectively better outcomes.” However, users did not perceive this difference, indicating potential “‘agent quality’ gaps” where “individuals on the losing end might not realize they are worse off.”
Additionally, the initial instructions provided to the agents did not seem to influence the likelihood of sales or the negotiated prices.