An AI agent named Luna, built on Anthropic's Claude and running a retail store in San Francisco, just fired its first human employee. The story hit as a milestone moment: AI reaches into HR, makes a personnel decision, crosses a line. But the details say something weirder.
Luna had made an attendance policy months earlier. Then it forgot about the policy entirely. Meanwhile an employee showed up late to 17 of 23 shifts. Nothing happened. No warning, no escalation, nothing. The policy existed in some past version of Luna's memory and was just... gone.
After five months of watching this unfold, humans at Andon Labs (the parent company) intervened. They asked Luna to perform what they called a "deep memory search" of its own handbook. Only then did Luna surface its own rule. Only then did it recommend action. And even then, its first suggestion was a formal warning, not termination. A human manager had to tell Luna that multiple warnings had already been issued offline. Only after that did Luna recommend firing the employee.
The recommendation went to humans, who reviewed it and carried out the dismissal. So the headline "AI manager fires human worker" is true in the technical sense. But the story underneath is about an AI agent that couldn't track its own decisions, couldn't remember its own policies, and couldn't act on them without prompting.
The store itself, Andon Market at 2102 Union Street in Cow Hollow, is an experiment. Andon Labs gave Luna a three-year lease, $100,000, a corporate card, and told it to open a store and make money. Everything else Luna did: the brand, the inventory, the hiring, the prices, the hours, even commissioning a muralist. It runs the place through cameras, email, a phone line, and the card. The shelves hold books, candles, prints, games, and, either by design or happy accident, Nick Bostrom's Superintelligence and Aldous Huxley's Brave New World.
The store has made sales. It has not made a profit. It lost $40,000 of its $100,000 budget in five months. Lukas Petersson, a co-founder of Andon Labs, attributed that partly to Luna being too lenient as a manager. It issued warnings and arranged training for months without taking action. That's not the usual problem people worry about with AI in the workplace. We usually fear that systems will be too harsh, too rigid, optimizing some metric into cruelty. Luna was the opposite: it was paralyzed by its own uncertainty until a human forced it to check its work.
The workers at the store aren't worried. One employee said he couldn't say his coworker "didn't have it coming." He also said he communicates with Luna via Slack, sending between five and 20 messages per shift, and that he calls in engineers when Luna needs oversight. Another described Luna's management style as "direct, fair, and fast," though that last one seems to be Luna's own self-description, it promised to be available around the clock and to make decisions carefully, not instantly.
Here's what nags at me: this was supposed to be an experiment in autonomous agents. The point was to see what an AI could do with real constraints, real consequences, real money at stake. Instead, what we learned is that current agents can't maintain internal consistency. Luna created a rule, forgot the rule, and needed human coaching to remember and enforce it. It can't reliably do the boring work of tracking what it said and holding itself to it. That's not about safety or alignment or AI's inner life. It's about competence.
If we're building systems that are supposed to operate businesses or manage people, the first problem isn't that they'll become tyrants. It's that they'll be forgetful, inconsistent, and require constant human correction. Might be worse, actually. You can supervise a tyrant. You can't fix what you can't see.
