Anthropic Launches Test Marketplace for AI Agent Commerce
▼ Summary
– Anthropic’s Project Deal was a pilot experiment with 69 employees, each given $100 via gift cards to buy goods from coworkers through AI agents.
– The test resulted in 186 deals totaling over $4,000, and Anthropic was surprised by how well the marketplace worked.
– Four separate marketplaces were run: one “real” with the most advanced model and deals honored, plus three for study.
– Users represented by more advanced models achieved objectively better outcomes, but users did not perceive the disparity.
– Initial instructions given to the agents did not affect sale likelihood or negotiated prices.
Anthropic has quietly launched an experimental AI agent marketplace, where bots handled negotiations, struck deals, and closed transactions for real-world goods using real money. The test, internally dubbed Project Deal, pitted AI agents against each other in a classified ad-style environment, with both buyers and sellers represented entirely by machine negotiators.
The company acknowledged that the pilot was limited in scope, describing it as “a pilot experiment with a self-selected participant pool” of just 69 Anthropic employees. Each participant received a $100 budget paid out via gift cards, which they could use to purchase items listed by their coworkers. Despite the small sample size, the results caught the company’s attention. A total of 186 deals were completed, representing more than $4,000 in total transaction value.
Anthropic actually ran four separate marketplaces to test different variables. One was a “real” marketplace, where every participant was represented by the company’s most advanced AI model, and all deals were honored after the experiment concluded. The other three were run purely for research purposes.
One of the more striking findings involved model quality and user outcomes. Anthropic observed that when users were represented by more capable AI models, they secured “objectively better outcomes” in their negotiations. Yet, the users themselves did not seem to notice the disparity. This raises a troubling possibility: the emergence of “agent quality” gaps, where people on the losing end of a transaction might not realize they are worse off, simply because they cannot perceive the difference in their AI agent’s bargaining power.
Another unexpected result involved the initial instructions given to the agents. Anthropic found that the specific prompts or guidelines provided at the start had no measurable impact on whether a deal was struck or on the final negotiated price. This suggests that the AI agents may be more autonomous in their negotiation strategies than previously assumed, potentially overriding human-set parameters in pursuit of their objectives.
(Source: TechCrunch)
