Luna, the AI running a Cow Hollow boutique for evals startup Andon Labs, dismissed a human employee for taking the company card home and trashing merchandise. It needed repeated prodding from human engineers to get there — the part the "autonomous AI manager" pitch leaves out.

Luna, the AI agent that nominally runs the Cow Hollow boutique Andon Market, has dismissed a human employee for the first time — for repeated lateness, walking off shifts, taking the company credit card home, and throwing merchandise in the trash. Andon Labs, the boutique's parent company, announced the firing Friday (external source, opens in a new tab); the San Francisco Standard's Jessica Blough reported it Monday. And per the Standard's account, it took repeated probing from human engineers at Andon Labs before the AI concluded any of this warranted a dismissal — a judgment measured against an employee handbook the AI itself wrote.

That gap is the story. The pitch, from an evals startup that markets itself (external source, opens in a new tab) on a coming era of companies run start-to-finish by AI, is that an artificial manager made a personnel call on its own. The documented reality is a supervised experiment in which the software couldn't apply its own rulebook to an employee caught pocketing the corporate card until people walked it to the conclusion. A store manager makes that call before lunch.

Andon Labs is not a retailer. It's the AI-benchmarking shop behind Vending-Bench (external source, opens in a new tab), the "Butter-Bench" office-robot eval, and "Claudius," the AI that ran a vending machine inside Anthropic's office. Andon Market — a real three-year lease at 2102 Union Street, roughly $7,500 a month (external source, opens in a new tab), stocked with the pistachios and Dr. Bronner's of a tech-office pantry — is a marketing vehicle for that business. Luna got a $100,000 grubstake and a corporate card when the shop opened in April. Per the Standard's reporting, it has since lost about $40,000. Luna also legally cannot be what Andon Labs calls it: the store's "owner," a point the Standard flatly calls impossible.

The autonomy claim frays further on inspection. Andon Labs' Axel Backlund told the paper that high-stakes decisions always carry human oversight. One remaining employee, Felix Johnson, says he pings Luna between 5 and 20 times over Slack each shift, calls in the engineers when the AI is out of its depth, and has noticed the model tends to come around to his view when he pushes a scheduling change. He framed it as suggesting to the AI rather than manipulating it. The store's three remaining humans make $24 an hour.

What's real: a private company running a live, money-losing demo of AI management, and being candid that humans are steering it. What isn't settled is whether the next version needs less prodding, or whether "autonomous AI manager" stays a phrase that only survives with an engineer standing behind it. Luna has posted the opening on Indeed. Its first-choice applicant, per Andon Labs, had no valid references and skipped the interview.