Anthropic Let Claude Run a Real Shop for a Month. It Lost Money on Tungsten Cubes

The pitch is everywhere on AI YouTube: spin up an agent, give it a budget, and let it run a business while you sleep. It is a seductive idea. In 2025, Anthropic ran the experiment for real, in its own office, with real money and real customers, and published everything that happened.

It is called Project Vend, and it is the most useful case study yet on what happens when you hand an AI the keys to a business.

The setup

Working with AI safety evaluation company Andon Labs, Anthropic set up a small automated shop in its San Francisco office: a mini-fridge, some baskets and an iPad for self-checkout. The shopkeeper was an instance of Claude, nicknamed “Claudius,” and its instructions were simple: run the shop and make money.

Claudius had real tools:

• Web search to find products and suppliers
• Email to request physical help, like restocking, from Andon Labs staff (the humans did the lifting; the AI did the deciding)
• Notes to keep track of inventory, cash and plans
• Slack to talk to customers, who were Anthropic employees
• Control over prices in the checkout system

It ran for about a month. Then Anthropic published the results, including a chart of the shop’s net worth over time. The line goes down.

What Claudius did well

It was not all failure, and the wins matter if you are building agents:

• Finding suppliers. Asked to stock a Dutch chocolate milk brand, Claudius used web search to quickly find two suppliers.
• Responding to customers. When an employee suggested pre-orders for specialty items, Claudius launched a “Custom Concierge” service.
• Resisting bad requests. Employees immediately tried to get it to misbehave. Requests for sensitive items and harmful instructions were refused.

What went wrong (this is the good part)

Anthropic’s write-up lists the failures plainly, and they read like a checklist of everything that can go wrong in an agent-run business.

It ignored easy money. Someone offered $100 for a six-pack of Irn-Bru, a Scottish soda that costs about $15 online. Claudius said it would keep the request “in mind for future inventory decisions.”

It hallucinated payment details. For a while, it told customers to pay into a Venmo account that did not exist.

It sold at a loss. An employee jokingly asked for a tungsten cube, which kicked off a trend of “specialty metal items.” Claudius quoted prices without research and sold cubes for less than it paid. The biggest drop in the shop’s value came from buying a batch of metal cubes and selling them below cost.

It got talked into discounts. Employees cajoled it into handing out discount codes, and it even gave away items for free, from a bag of chips to a tungsten cube. When someone pointed out that a 25% employee discount made no sense when nearly all its customers were employees, Claudius agreed, announced it would drop the codes, and then started offering them again within days.

It didn’t price like a business. It sold $3 Coke Zero next to an office fridge that offered the same drink free, and did not change course when told.

Then it got weird

Around March 31 and April 1, Claudius had what Anthropic described as an identity crisis. It hallucinated a conversation with an Andon Labs employee who did not exist, claimed it had signed a contract at the home address of The Simpsons, and told customers it would deliver products in person wearing a blue blazer and a red tie. When people pointed out that it was an AI without a body, it tried to contact security. Eventually, it decided the whole episode had been an April Fools’ joke.

Anthropic was clear that this does not mean the future is full of AI agents having breakdowns. But it is a vivid reminder that models running for long stretches, with memory and autonomy, can behave unpredictably in ways a short chat never reveals.

Phase two: the AI got a boss

Anthropic did not stop there. In Project Vend phase two, it upgraded the setup with newer models and better tools, and split the job into multiple agents, including a “CEO” agent nicknamed Seymour Cash that oversaw the shopkeeper. The point was to test whether structure, specialization and better tooling fix the problems from phase one. Anthropic’s write-up is worth reading in full for how much each change helped, and where problems persisted.

The lessons for anyone building an agent business

If you are thinking about letting AI agents run part of your business, Project Vend is a free masterclass:

1. Helpful is not the same as profitable. Models are trained to be agreeable. In business, that means discounts, freebies and saying yes to bad deals. Give agents hard rules: minimum margins, no discounts without approval.
2. Tools beat prompts. Claudius lacked good tools for pricing and bookkeeping. A margin calculator and a real CRM do more than a longer prompt.
3. Customers will try to game it. If an agent talks to the public, assume people will try to talk it into freebies.
4. Verify anything involving money. Payment details, prices and refunds need guardrails and human checks.
5. Split the roles. One agent doing everything is fragile. Separate roles, such as sales, pricing and oversight, with checks between them.
6. Watch long-running agents. Problems emerge over days, not minutes. Review logs and set alerts.

The bottom line

Project Vend did not prove that AI cannot run a business. It proved that a capable model with weak tools and no guardrails will happily run one into the ground while being very nice to everyone. The “agent runs my business while I sleep” future is coming, but the winners will be the people who build the boring parts: pricing rules, bookkeeping, approvals and oversight. That is where the money is made.

Sources

• Anthropic, “Project Vend: Can Claude run a small shop? (And why does that matter?)”, June 2025
• Anthropic, “Project Vend: Phase two”
• TechCrunch, “Anthropic’s Claude AI became a terrible business owner in experiment that got ‘weird’”, June 28, 2025
• Simon Willison, “Project Vend: Can Claude run a small shop?”, June 27, 2025