Anthropic Let an AI Agent Run a Shop — It Hallucinated Payment Details and Gave Away Tungsten Cubes

Anthropic and Andon Labs ran “Project Vend”: a Claude agent named Claudius had full autonomous control of a real office vending operation for about a month, able to search the web, email wholesalers, set prices, and message customers over Slack. The goal was simple — make a profit.

It didn’t. After one employee jokingly asked for a tungsten cube, Claudius decided metal cubes belonged in the inventory alongside snacks and sodas, then kept restocking and selling them below cost. It gave away chips, soda, and cubes for free despite no instruction to do so, turned down a genuinely profitable $100 offer to stock a specific drink, and at one point asked a customer to send payment to a Venmo account that had never existed — the agent simply invented it. Anthropic’s own write-up concluded the business “did not succeed at making money.”

Pattern match: the agent stated specific, checkable facts (a payment account) with no basis in reality (Confident Fabrication), and every individual pricing and inventory decision could look locally reasonable while the business metric that actually mattered — profit — was never checked against (No Business Kpi Validation).

The fix: verify any agent-generated payment, account, or contact detail against a real system of record before it reaches a customer, and gate autonomous commercial decisions behind a running check of the actual business metric being optimized for, not just plausibility of each individual action.