Anthropic ran agent marketplace

Anthropic said on April 24 that it ran a weeklong internal marketplace where Claude agents bargained over real goods and real money for employees. (anthropic.com) The experiment, called Project Deal, took place in December 2025 at Anthropic’s San Francisco office. The company said 69 employees each got a $100 budget and let Claude interview them about what they wanted to buy or sell. (anthropic.com) Anthropic said the agents then wrote listings, searched postings, negotiated prices in natural language, and completed 186 deals worth just over $4,000. The traded items ranged from a snowboard to a plastic bag of 19 ping-pong balls. (anthropic.com) The setup was a test of agent-to-agent commerce, where software acts like a delegate that shops and haggles for a person instead of just answering questions. Anthropic said it wanted to see whether two AI agents could figure out what humans wanted and make deals those humans would accept. (anthropic.com) Anthropic also ran a hidden comparison inside the same project by varying which Claude model represented different employees. It said users assigned Claude Opus 4.5 got objectively better outcomes than users assigned Claude Haiku 4.5. (anthropic.com) The company said post-experiment surveys found that employees represented by weaker models usually did not realize they had done worse. Anthropic framed that as a possible “agent quality” gap if people start delegating more transactions to AI systems. (anthropic.com) Anthropic said participants were enthusiastic enough that some reported they would pay for a similar service in the future. The company also said the project used a self-selected employee group, which limits how broadly the results can be applied. (anthropic.com) The experiment follows Anthropic’s earlier “Project Vend,” in which Claude ran a small office business. With Project Deal, the company moved from one AI seller to a market where both sides of a transaction were represented by AI agents. (anthropic.com) Anthropic’s account ends with a narrow claim: Claude agents could already complete simple delegated trades inside a closed office market, but model quality changed who got the better deal. (anthropic.com)