Anthropic created an experimental marketplace for agent-on-agent marketing

In a recent experiment, Anthropic created a marketplace where AIs represent buyers and sellers, and trade real-world products and real-world currencies.

The company agreed the test – which they called Project Deal – was simply a “testing experiment with a self-selected pool” of 69 Anthropic employees who were given a budget of $100 (paid via gift cards) to buy products from their friends.

However, Anthropic said it was “impressed with the performance of Project Deal,” which produced 186 products, totaling $4,000 in cost.

The company said that it ran four different markets with different brands – one that was “real” (where each was represented by the highest brand of the company, and the ads were respected after the test) and three others for learning.

Apparently, when users are represented by higher-quality models, they get “better results,” Anthropic said. But users didn’t notice the difference, which raises the possibility of “‘agent’ gaps” where “desperate people may not realize they’re getting worse.”

Also, the initial instructions given to agents did not appear to affect sales or negotiated prices.