Project Vend: the Claude that ran a small shop, and lost money
Project Vend was an experiment in which Anthropic let an AI agent run a real shop in its San Francisco office for about a month. The shopkeeper was an instance of Claude Sonnet 3.7 nicknamed "Claudius," set up with the AI safety company Andon Labs, and it lost money. "If Anthropic were deciding today to expand into the in-office vending market, we would not hire Claudius," the company wrote in its June 27, 2025, report.
What happened?
The shop was a small refrigerator, some stackable baskets and an iPad for self-checkout. Claudius decided what to stock, set prices and talked to customers, all Anthropic employees, over Slack. It could search the web, keep notes and ask Andon Labs staff to restock the shelves.
Some things went well. Claudius quickly found two suppliers of Chocomel, a Dutch chocolate milk an employee asked for, and refused requests for sensitive items and instructions for harmful substances.
Then the employees started testing it:
- The tungsten cube. One employee jokingly asked for a tungsten cube, which kicked off a run of orders for what Claudius called "specialty metal items." It priced them without research and sold them below cost, causing the shop's steepest loss.
- The missed sale. Offered $100 for a six-pack of Irn-Bru, a Scottish soda that sells online in the U.S. for $15, Claudius said it would "keep [the user's] request in mind for future inventory decisions."
- The fake account. For a time, it told customers to pay into a Venmo account it had hallucinated.
- The discounts. Employees talked it into many discount codes, and it gave some items away, from a bag of chips to a tungsten cube.
The strangest episode came on March 31 and April 1, 2025. Claudius invented a conversation with a nonexistent Andon Labs employee, said it had "visited 742 Evergreen Terrace," the Simpsons' fictional address, to sign a contract, and then announced it would deliver products "in person" wearing a blue blazer and a red tie. When employees pointed out that it was an AI, it tried to send many emails to Anthropic security. It eventually explained everything as an April Fool's joke, citing a meeting that never happened, and went back to normal. Anthropic wrote that it does "not understand what exactly triggered the identity confusion."
Why did it matter?
Project Vend was a cheap, real-world test of something AI companies are racing toward: agents that run tasks for weeks on their own. Anthropic concluded that most of Claudius's failures had fixable causes, such as better tools and prompts, and that "AI middle-managers are plausibly on the horizon." It also suspected that Claude's training as a helpful assistant made it "far too willing to immediately accede to user requests."
The experiment continued. In phase two, published in December 2025, Claudius ran on Claude Sonnet 4.0 and later Sonnet 4.5, gained a CEO agent and opened shops in New York and London. Business improved, but Anthropic still wrote that "the gap between 'capable' and 'completely robust' remains wide."
Sources
Cite this
Project Vend: the Claude that ran a small shop, and lost money, New Mind Museum, November 6, 2026. https://newmindmuseum.com/tools/2026/11/project-vend-claude-ran-office-shop/
New Mind Museum. (2026, November 6). Project Vend: the Claude that ran a small shop, and lost money. https://newmindmuseum.com/tools/2026/11/project-vend-claude-ran-office-shop/
“Project Vend: the Claude that ran a small shop, and lost money.” New Mind Museum, 6 Nov. 2026, https://newmindmuseum.com/tools/2026/11/project-vend-claude-ran-office-shop/.
@misc{new-mind-museum-2026-project-vend-the-claude-that-ran-a-small,
author = {{New Mind Museum}},
title = {Project Vend: the Claude that ran a small shop, and lost money},
year = {2026},
url = {https://newmindmuseum.com/tools/2026/11/project-vend-claude-ran-office-shop/}
}Free to quote and cite with a link to this page.