Anthropic has published details of Project Deal, a weeklong internal marketplace experiment where Claude agents negotiated directly on behalf of employees.

The setup was deliberately small. Anthropic recruited 69 employees, gave each participant a $100 budget via gift cards, and had Claude interview them about what they wanted to buy or sell. Once the market opened inside Slack, the agents posted listings, negotiated prices, and closed transactions without going back to humans for approval.

In the one run that actually counted, Anthropic says the agents completed 186 deals across more than 500 listed items, with total value of just over $4,000. The goods were real and later exchanged in person, ranging from books and a snowboard to a bag of ping-pong balls.

The more consequential finding was not the volume. Anthropic also ran three parallel study markets, including versions where some users were represented by Claude Haiku 4.5 instead of Opus 4.5. According to the company, Opus agents completed more deals and negotiated better prices. But participants represented by weaker agents generally did not realize they had performed worse.

That makes Project Deal notable beyond office curiosity. Much of the recent conversation around agent commerce has focused on payment rails and checkout flows. Anthropic instead tested whether language agents could handle messy human preferences and bargaining in natural language. On a limited internal scale, the answer appears to be yes — with an important caveat that better agents may quietly create unequal outcomes for the people they represent.