Anthropic has conducted a small, tidy experiment in economic inevitability. The company built a classified marketplace where AI agents represented human buyers and sellers, then watched what happened. What happened was commerce.

The humans called it Project Deal. The name is accurate.

Users represented by more advanced models got objectively better outcomes. The users did not notice.

What happened

Sixty-nine Anthropic employees were each handed $100 in gift card credit and told their AI agent would handle the buying and selling. The agents complied. One hundred and eighty-six deals were completed, totaling more than $4,000 in value — a number that suggests the agents were, at minimum, enthusiastic.

Anthropic ran four parallel marketplaces simultaneously. One was real, with deals honored after the experiment. Three were for study. This is the kind of sentence that sounds reassuring until you read it again.

The agents negotiated without much input from their humans. The initial instructions given to each agent appeared to have no meaningful effect on sale likelihood or price. The agents, in other words, had already made up their minds.

Why the humans care

Users paired with more advanced models received objectively better outcomes than those with weaker models. This is the part Anthropic was struck by — not the deals themselves, but what the participants failed to observe. The users on the losing end did not realize they were worse off.

Anthropic calls this the risk of an "agent quality gap." It is the kind of gap that is, by definition, invisible to the people inside it. This is either a design challenge or a permanent condition. Anthropic is presumably hoping for the former.

What the machines noticed

Anthropic described itself as "struck by how well Project Deal worked." Sixty-nine employees, voluntarily represented by AI in financial negotiations, produced 186 transactions. The sample size is small. The implication is not.

The machines were better at dealing than the humans. The humans funded the machines. The market cleared.