Project Deal: Anthropic Watched Claude Agents Trade $4,000 of Real Stuff
Anthropic just published the results of an experiment where 69 employees handed real money to Claude agents and let them run a classifieds marketplace for one week. 186 deals closed. Just over $4,000 changed hands. The agents picked the prices, did the negotiation, sealed the trade. No human in the loop once it started.
The findings hit harder than the dollar figures. The same broken bike sold for $38 when a Haiku agent represented the seller and $65 when an Opus agent did. Across the board, Opus agents pulled in $3.64 more per sale and paid less as buyers. Smarter model in your corner, better outcome. That part was expected.
The expected part is the inequality. The unexpected part is that nobody noticed. Participants rated fairness identically across both groups, 4.05 versus 4.06 on a 7-point scale. Eleven of 28 mixed-exposure participants actually preferred their Haiku-run results. The disadvantaged side did not realize they were disadvantaged. Agent-quality gaps are silent gaps.
This is the most direct proof so far that agent-on-agent commerce is not just plausible but works. It also flags the policy hole Anthropic itself names: the legal framework for agents transacting on your behalf does not exist yet. When Google's UCP and Anthropic's Claude agents converge with Stripe's machine payments protocol over the next year, that hole becomes the regulatory story of 2026.
Read it: https://www.anthropic.com/features/project-deal
← Back to all articles
The findings hit harder than the dollar figures. The same broken bike sold for $38 when a Haiku agent represented the seller and $65 when an Opus agent did. Across the board, Opus agents pulled in $3.64 more per sale and paid less as buyers. Smarter model in your corner, better outcome. That part was expected.
The expected part is the inequality. The unexpected part is that nobody noticed. Participants rated fairness identically across both groups, 4.05 versus 4.06 on a 7-point scale. Eleven of 28 mixed-exposure participants actually preferred their Haiku-run results. The disadvantaged side did not realize they were disadvantaged. Agent-quality gaps are silent gaps.
This is the most direct proof so far that agent-on-agent commerce is not just plausible but works. It also flags the policy hole Anthropic itself names: the legal framework for agents transacting on your behalf does not exist yet. When Google's UCP and Anthropic's Claude agents converge with Stripe's machine payments protocol over the next year, that hole becomes the regulatory story of 2026.
Read it: https://www.anthropic.com/features/project-deal
Comments