Anthropic's Project Deal involved 69 employees

Anthropic said on April 24 that it had run an internal marketplace experiment in which AI agents negotiated purchases and sales on behalf of employees. The company said 69 workers in its San Francisco office took part, each with a $100 budget, and the agents completed 186 deals worth just over $4,000. TechTimes cited the experiment in a May 17 article about the infrastructure and legal gaps around autonomous agent payments. Anthropic described the project as a pilot with a self-selected participant pool. ### How did Anthropic set up the marketplace? Anthropic said the experiment took place over one week, with a classified marketplace built for employees in its San Francisco office. The company said the setup worked like Craigslist, except AI models handled the deals for both sides. In December 2025, according to Anthropic’s write-up, Claude interviewed employees about items they might sell and things they might want to buy. The company said it gave each participant’s agent $100 to spend, then let those agents post listings, negotiate and make agreements. ### What actually changed hands? Anthropic said the goods were real and the transactions were honored after the experiment ended. The company said employees later brought in the physical items and exchanged them in person. The list of items included a snowboard and a plastic bag of ping-pong balls, according to Anthropic’s account. The company said the agents had haggled over those goods through their AI “avatars” before the human handoff. ### What did the numbers show? Anthropic said the marketplace produced 186 deals with a total transaction value of just over $4,000. TechCrunch, citing Anthropic’s experiment, reported that participants received the $100 budgets through gift cards. The company said participants responded positively in post-experiment feedback and some said they would pay for a similar service in the future. Anthropic also said the project was limited by its small, self-selected group, a caveat it included in the published write-up. ### Why did Anthropic run a second, hidden test? Anthropic said it ran a parallel experiment in secret to test whether model quality affected bargaining outcomes. The company said it compared its then-frontier Claude Opus 4.5 model with its smaller Claude Haiku 4.5 model. The result, Anthropic said, was that people represented by stronger models got “objectively better outcomes.” The company also said post-experiment surveys showed participants using weaker models did not notice they had been disadvantaged. ### Where does this fit in the broader agent-commerce push? TechTimes said on May 17 that Anthropic’s Project Deal offered one example of agent-to-agent commerce as payment infrastructure expands. The same article said Amazon Web Services launched Bedrock AgentCore Payments on May 7, using stablecoin micropayments built with Coinbase and Stripe. TechTimes also reported that OpenAI had pulled its Instant Checkout feature in March 2026 after limited merchant adoption. In that framing, Project Deal was presented as evidence that autonomous agents can negotiate and complete transactions, even as consumer protections and dispute rules lag behind the technology. ### Where can readers find the original details? Anthropic published the Project Deal write-up on its website on April 24, 2026, under the headline “Project Deal: our Claude-run marketplace experiment.” TechTimes referred to the experiment in its May 17, 2026 report on AI agents buying, hiring and paying other agents. TechCrunch also summarized the experiment on April 25, reporting that Anthropic had run four marketplaces, including one in which deals were actually honored after the test. Anthropic’s published page remains the primary source for the participant count, the $100 budgets, the 186 completed deals and the company’s account of what happened next.