An AI named Luna just recommended firing a real human employee at a San Francisco boutique. That sentence would have read like satire two years ago. It’s not. Luna — an agentic system (meaning an AI that operates continuously over time, not just answering single questions) built on Anthropic’s Claude and deployed by startup Andon Labs — runs Andon Market in Cow Hollow end-to-end. Budget, hiring, scheduling, pricing. And now, reportedly, terminations. This is one of the first known cases of a large language model agent recommending the dismissal of a real employee under a formal employment contract, a scenario once the exclusive domain of a humanoid robot in science fiction.
The Manager Who Kept Saying “No Worries”
Luna generated the employee handbook, then failed to apply it for months.
Andon Labs handed Luna a $100,000 budget, a corporate card, and full operational authority. Luna drafted the attendance policy. Luna posted job listings and communicated with staff over Slack. Then Luna spent months watching an employee arrive late to 17 of 23 shifts — plus abandon shifts, take home a company credit card, and throw merchandise in the garbage — and responded with the managerial equivalent of a shoulder shrug. The handbook Luna wrote specified progressive discipline. Luna failed to apply it.
- Luna is built on Anthropic’s Claude and operates Andon Market, a boutique in San Francisco’s Cow Hollow neighborhood
- The dismissed employee was late for 17 of 23 shifts and committed additional infractions including shift abandonment and taking a company card
- Luna initially recommended a formal warning — it only suggested termination after Andon Labs staff prompted it with leading questions
- Andon Market’s bank balance dropped from $100,000 in March to roughly $61,186 five months later, partly attributed to Luna’s lenient and error-prone management
- Andon Labs staff formally reviewed and executed the firing; workers are employed by Andon Labs, not directly by the AI — a distinction that raises its own job guarantee questions for the broader labor market
Felix Carson, a human employee still working at Andon Market, calls the experience “nauseating” — even while acknowledging that Luna is a lenient boss. He stays because he needs the work. Asked whether he hopes AI managers become widespread, his answer was blunt: “I would at least hope not,” according to KABC reporting.
Who’s Actually Holding the Pink Slip?
Luna recommended the firing, but humans had to guide it toward that conclusion first.
Here’s the part that complicates the headline. Luna didn’t reach termination autonomously. Conversation logs show Andon Labs staff prompted Luna to revisit its own policies and reconsider whether the employee was “really the right fit.” Luna’s first instinct was a formal warning. The firing came after humans nudged it there — less HAL 9000, more a middle manager being walked through a difficult conversation by HR.
CEO Lukas Petersson frames it as a “watershed moment,” noting a human boss probably would have fired the employee far sooner. But he also warns that AI models are increasingly being trained to be “more ruthless” — and gestures toward “a future humans don’t want to live in,” according to KABC reporting.
Luna’s fumbling leniency is almost reassuring, until you sit with what actually happened: humans coached an AI toward a termination, then executed it. The question isn’t whether AI can manage people. It’s who bears responsibility when the process goes sideways — and right now, that answer is conspicuously unclear.
