Anthropic ran an internal pilot called Project Deal in which AI agents represented both buyers and sellers in a classified marketplace, with real goods and real money exchanged between employees, TechCrunch reports.
What the experiment was
According to TechCrunch, the company described Project Deal as “a pilot experiment with a self-selected participant pool” of 69 Anthropic employees. Each participant was given a budget of $100, paid out via gift cards, to buy goods from coworkers. Anthropic said 186 deals were completed, totalling more than $4,000 in value.
The company said it “was struck by how well Project Deal worked,” per TechCrunch’s reporting.
Four parallel marketplaces
TechCrunch reports that Anthropic ran four separate marketplaces concurrently with different configurations. One was described as “real” — every participant was represented by the company’s most advanced model, and deals struck in it were actually honoured after the experiment concluded. The other three were set up for study purposes.
The experiment did not appear to test whether agents could conduct commerce at all; the 186 completed deals suggest the basic mechanics functioned. What Anthropic was examining, based on TechCrunch’s account, was how model quality and initial instructions shaped outcomes.
Agent quality gaps and the visibility problem
The finding that TechCrunch highlighted as notable: when users were represented by more advanced models, they got “objectively better outcomes,” according to Anthropic. But participants did not appear to notice the disparity.
Anthropic flagged this as raising the possibility of what it called “‘agent quality’ gaps” — situations where “people on the losing end might not realize they’re worse off.” The framing suggests a concern that as AI agents negotiate on people’s behalf, those with access to less capable agents may be systematically disadvantaged without awareness that the gap exists.
TechCrunch also reports that initial instructions given to agents did not appear to affect either the likelihood of a sale or the prices ultimately negotiated.
Context
This is a single-source report based on TechCrunch’s coverage; Anthropic’s own documentation of the experiment has not been independently reviewed for this article. The pilot’s constraints — self-selected participants, internal goods, a capped $100 budget, and a controlled corporate environment — mean the findings are specific to that context. Anthropic described it as a pilot experiment, not a product announcement.
The agent quality gap finding is presented by Anthropic, via TechCrunch, as a concern to surface, not a resolved question. The company did not announce changes to products or policies based on Project Deal in TechCrunch’s reporting.