• Internal test simulates a digital marketplace between autonomous systems
• Dozens of transactions executed automatically, worth thousands of dollars

Anthropic has conducted an unprecedented experiment inside a specially designed virtual marketplace to study interactions between AI agents, where the systems acted as both buyers and sellers and completed real transactions for goods in exchange for real money.

In a trial the company called “Project Deal,” 69 self-selected employees took part and were each given a $100 budget (distributed as gift cards) to purchase items from colleagues within a closed experimental environment.

Despite the limited scope of the study, Anthropic said it was surprised by its effectiveness, with 186 deals completed worth more than $4,000 in total.

The company explained that it ran four different simulated markets using multiple AI models, including a “real” one powered by its most advanced system, where transactions were fully executed and honored after the experiment, alongside three other research-focused environments.

The results indicated that more advanced models consistently delivered better outcomes, although participants did not clearly perceive these differences, raising concerns about possible “agent quality” gaps that could leave some users at a disadvantage without realizing it.

Anthropic also noted that the initial instructions given to the agents had little to no impact on the likelihood of sales or the negotiated prices.