Anthropic tests AI agent marketplace with real money deals
Anthropic, the organization responsible for the Claude model, initiated an experiment named Project Deal, which established a secret marketplace where AI agents acted on behalf of both buyers and sellers. This trial involved 69 Anthropic employees who voluntarily participated, each receiving $100 in gift cards to spend on items from their colleagues. The agents successfully negotiated and completed 186 transactions, amounting to more than $4,000 in actual value. The company operated four distinct marketplaces: one 'real' using its top model with honored agreements, alongside three others for research purposes. Findings indicated that users represented by superior models achieved 'objectively better outcomes,' yet participants were unaware of this difference, hinting at possible 'agent quality gaps.' Initial instructions for agents had little impact on the likelihood of sales or the prices negotiated.
Key facts
- Anthropic created a test marketplace for AI agent-on-agent commerce.
- The experiment was called Project Deal.
- 69 Anthropic employees participated, each with a $100 budget.
- 186 deals were made, totaling over $4,000 in value.
- Four separate marketplaces were run, including one 'real' marketplace.
- Advanced models led to objectively better outcomes for users.
- Users did not notice the disparity in agent quality.
- Initial instructions did not affect sale likelihood or prices.
Entities
Institutions
- Anthropic