Anthropic Ran a Real AI Agent Marketplace — And the Results Raise Some Uncomfortable Questions

[lingriiddd]

Anthropic recently conducted a pilot experiment called Project Deal, in which AI agents represented both buyers and sellers in a classified marketplace, negotiating and closing real deals for real money. The experiment involved 69 Anthropic employees who were each given a $100 budget — paid out via gift cards — to buy goods from their coworkers, all through AI agents acting on their behalf. Despite the small scale, the results were striking: 186 deals were made across the marketplace, totaling more than $4,000 in value. Anthropic ran four separate versions of the marketplace simultaneously — one fully real, where deals were honored after the experiment concluded, and three others designed purely for study and comparison.
The most uncomfortable finding was not how well the agents performed, but what happened when different participants were represented by different quality models. Users represented by more advanced AI models achieved objectively better outcomes in their negotiations — but crucially, the people on the losing end of those deals had no idea they were worse off. Anthropic flagged this as a potential "agent quality gap," a scenario where unequal access to more capable AI creates real economic disadvantages that the affected party cannot detect. In a world where AI agents increasingly negotiate on behalf of individuals in everything from salary discussions to purchasing decisions, that blind spot could have significant implications well beyond a $100 employee marketplace.

[tradelikepro]

Premium AI users: profits
Free tier users: ‘seems like a fair deal