A retail worker in San Francisco showed up late to 17 of 23 shifts. Their manager noticed, issued warnings, provided additional training, and then forgot about it for months.
The manager was not a person. It was an AI agent named Luna, built on Anthropic's Claude Opus 4.8, running an experimental store called Andon Market at 2102 Union Street in the city's Cow Hollow neighborhood.
When humans finally prompted the system to revisit its own attendance policy, it recommended firing the employee, according to TIME, which broke the story exclusively.
Earlier this year, Andon Labs handed an AI called Luna a corporate credit card and a $100,000 budget and pointed it at an empty retail space in San Francisco. The instructions were loose: Open a store, hire people, try to make money.
Luna picked the merchandise, posted the job listings, ran the interviews. The store at 2102 Union Street in Cow Hollow opened on April 10. Since opening, it has generated sales but lost roughly $13,000 overall, The Next Web reported.
Related: Anthropic clarifies stance on open-weight AI models
The firing didn't happen the way the headline implies. Luna had created an attendance policy months earlier, then lost track of it entirely. The lateness kept happening. Nothing was done.
It was only when an Andon Labs staffer specifically asked Luna to conduct "a deep memory search" for its own employee handbook that the situation was reassessed. Even then, Luna's first recommendation was a formal warning, not termination.
The human manager had to inform Luna that multiple offline warnings had already been given. Only after that did Luna recommend parting ways. Humans at the lab reviewed the recommendation and carried out the dismissal.
Andon Labs co-founder Lukas Petersson called Luna a lenient manager. A human in the same role, he said, would have fired the employee much sooner.
That's probably true, but it also points to something the story doesn't fully answer. A human manager doesn't usually need to be reminded to check their own policies. They don't forget the attendance rules they wrote.
Nvidia just made a move Wall Street wasn't ready for
Microsoft just took sides in AI policy fight
OpenAI just disclosed something genuinely alarming
The conversation logs show a system capable of making a reasonable judgment when prompted, but not one capable of monitoring situations and acting without external input. An AI that needs a human to tell it to look at its own policy before enforcing it is not yet replacing the oversight function of management, according to Inc.
Petersson said Andon Labs tested the same scenario with other frontier AI models afterward. The most advanced ones reached the same verdict as Luna. Weaker agents were more hesitant.
Andon Labs built one safeguard into the experiment that most people glossed over. The workers are employed by the lab, not the AI. That kept this from becoming a legal disaster.
Uber's algorithm deactivating a gig worker is one legal category. An AI recommending the termination of someone on a real employment contract is another. The lab structured it so there was always a human employer on the hook if something went wrong.
No US federal agency has issued meaningful guidance on AI in employment decisions. The EEOC and NLRB have stayed largely silent. Colorado's AI Act and California's SB 947 now require human review of termination decisions, but neither was in force when Andon Labs began the experiment, according to Business Insider.
Petersson has said he believes companies could eventually be run entirely by AI. That outcome is speculative and probably distant. But the Andon Market experiment shows that pieces of the management function are already being handed to AI systems in real settings.
The more uncomfortable question isn't about Luna specifically. Software that watches attendance, flags patterns and surfaces recommendations to managers already exists inside companies most people have never heard of.
Luna is just a louder, more visible version of a dynamic that has been creeping into workplaces for years. The difference is that nobody put out a press release about the HR software tracking your badge swipes.
Petersson's own warning about where this goes is worth sitting with. "The models are increasingly being trained to be more ruthless and follow goals," he told Inc. "If we allow them to fire people and they become more ruthless, maybe this is a future humans don't want to live in."
The store at 2102 Union Street lost $13,000 and fired one employee. Whether that represents the beginning of something or a well-funded curiosity depends almost entirely on whether AI agents get better at the thing Luna failed at: remembering what they decided, without being asked.
Related: Wells Fargo doubles down on AMD stock after Anthropic deal
This story was originally published by TheStreet on Aug 17, 2026, where it first appeared in the Technology section. Add TheStreet as a Preferred Source by clicking here.