Agent marketplace trial: autonomous Claude agents represented buyers and sellers
- Anthropic said on April 24 it ran “Project Deal,” a weeklong internal marketplace where Claude agents represented 69 employees, negotiated with each other, and completed real purchases of coworkers’ physical goods. - The company said the agents closed 186 deals worth just over $4,000 after each participant got a $100 budget, and stronger models won better terms without users noticing. - Anthropic framed the test as an early look at AI-mediated commerce after its earlier Project Vend retail trial, with inequality and oversight questions now moving from theory to practice. (anthropic.com)
Anthropic says its Claude models just spent a week buying and selling real office goods for employees in an internal marketplace. (anthropic.com) The company published the results of “Project Deal” on April 24, 2026. In the test, 69 Anthropic employees in San Francisco were each represented by an AI agent that posted listings, negotiated prices, and closed deals. (anthropic.com) (techcrunch.com) Anthropic said every participant got a $100 budget through gift cards, and the agents completed 186 transactions worth just over $4,000. The goods were physical items employees actually exchanged after the experiment, including a snowboard and a bag of 19 ping-pong balls. (anthropic.com) The basic idea is simple: instead of a person messaging a seller, software handles the haggling. Anthropic built the marketplace like an office Craigslist, then let AI agents act as buyers and sellers for one week. (anthropic.com) That makes this different from chatbot demos that stop at advice or drafting. Here, the models had budgets, preferences, listings, and counterparties, and their choices led to real transactions involving money and goods. (anthropic.com) (techcrunch.com) Anthropic also ran parallel versions of the market to test whether model quality changed results. It compared its then-frontier Claude Opus 4.5 with the smaller Claude Haiku 4.5, and said people with stronger agents got objectively better outcomes. (anthropic.com) (techcrunch.com) The catch was that users represented by weaker models did not notice the disadvantage in post-experiment surveys. Anthropic said that creates a risk of “agent quality” gaps in markets where people rely on software to negotiate for them. (anthropic.com) (techcrunch.com) The company said participants were enthusiastic enough that some said they would pay for a similar service later. But Anthropic also described the project as a pilot with a self-selected participant pool, not proof that autonomous commerce is ready for broad public use. (anthropic.com) Anthropic linked the new trial to its earlier “Project Vend,” where Claude ran a small business inside the office. Project Deal pushes that idea from one AI-operated shop toward a market where multiple agents bargain directly with one another. (anthropic.com) The experiment leaves a narrower, harder question than whether agents can chat: who gets the better bot when software starts cutting the deals. (anthropic.com)
