Anthropic ran agent marketplace
- Anthropic said April 24 it ran “Project Deal,” a weeklong internal marketplace where Claude agents bought and sold coworkers’ physical goods in San Francisco. - The pilot involved 69 employees with $100 each; agents closed 186 deals worth more than $4,000, from a snowboard to 19 ping-pong balls. - Anthropic also found stronger models won better outcomes, while weaker-model users often missed the gap. (anthropic.com)
Anthropic said on April 24 that it ran an internal marketplace where Claude agents negotiated real purchases and sales for employees. (anthropic.com) The project, called Project Deal, ran for one week in Anthropic’s San Francisco office and covered physical goods employees already owned. Anthropic said the setup worked like a classifieds board, except the haggling was done by AI agents. (anthropic.com) In December 2025, Claude interviewed employees about what they might want to sell and what they hoped to buy. Anthropic gave each participant’s agent a $100 budget through gift cards before the marketplace opened. (anthropic.com) (techcrunch.com) Anthropic said the agents completed 186 deals worth just over $4,000 by the end of the experiment. The items ranged from a snowboard to a plastic bag containing 19 ping-pong balls. (anthropic.com) The basic idea is agent commerce: software acts like a personal shopper and seller, taking instructions, comparing options, bargaining, and closing a transaction. Anthropic said it wanted to test what happens when both sides of a marketplace are represented by AI instead of humans. (anthropic.com) Anthropic also ran parallel versions of the market to study model quality, including a comparison between Claude Opus 4.5 and Claude Haiku 4.5. The company said people represented by the stronger model got objectively better outcomes. (anthropic.com) (techcrunch.com) In post-experiment surveys, Anthropic said participants using weaker models generally did not notice that they were getting worse deals. The company said that creates a risk of “agent quality” gaps that users may not detect on their own. (anthropic.com) (techcrunch.com) Anthropic called Project Deal a pilot with a self-selected pool of 69 employees, not a public product launch or a broad market test. The company also said participants were enthusiastic enough to report they would pay for a similar service in the future. (anthropic.com) The experiment follows Anthropic’s earlier “Project Vend,” which tested whether Claude could run a small office shop. Project Deal moved that work from stocking shelves to representing buyers and sellers in direct negotiations. (anthropic.com) The closing point from Anthropic was narrower than the headline: the company said agent marketplaces already look feasible in a small office setting, but model quality changes who wins. (anthropic.com)