Anthropic Tests AI Agent Marketplace, 186 Deals Totaling Over $4,000

AI safety company Anthropic has revealed details of an internal experiment called “Project Deal,” in which it built a classified marketplace where AI agents represented both buyers and sellers, successfully executing real transactions for real goods and real money.

Experiment Design

Anthropic recruited 69 employees for the pilot, each given a $100 budget (paid out in gift cards) to purchase items from coworkers. The company actually ran four separate marketplaces with different model configurations — one “real” market where all participants were represented by the company’s most advanced model, with deals honored after the experiment concluded.

Key Findings

The experiment exceeded expectations, completing 186 deals with a total value exceeding $4,000. Anthropic stated it was “struck by how well Project Deal worked.”

More notably, users represented by more advanced AI models achieved “objectively better outcomes” in negotiations. However, participants did not seem to notice this disparity, raising concerns about “agent quality gaps” — where people on the losing end of a negotiation might not realize they are worse off.

Additionally, the researchers found that the initial instructions given to the agents did not significantly affect the likelihood of sales or negotiated prices, suggesting that the inherent capability of the agent may matter more than its preset strategies.

Industry Implications

The experiment reveals the potential of AI agents in automated commerce while raising new ethical questions. When AI agents represent humans in transactions, differences in “agent quality” could lead to information asymmetry that undermines market fairness.

As AI agent technology matures, ensuring equitable outcomes across different quality tiers of agents will become an important challenge the industry must address.

Source: TechCrunch