Written By | Torch Media Africa
Inside the San Francisco corner store where an AI named Luna hired, managed, and finally sacked a flesh-and-blood employee — and why Silicon Valley says this is only the beginning
Search
SAN FRANCISCO
There's a boutique on Union Street in the leafy Cow Hollow neighborhood that looks like any other trendy San Francisco storefront: exposed brick, a hand-painted mural out front, artfully arranged merchandise on reclaimed-wood shelving. Customers wander in, browse, chat with the staff at the register. Nothing about it screams "the future of work just walked through that door."
But it did. Because the person who designed the mural, set the prices, wrote the employee handbook, hired the staff — and, this month, fired one of them — was never a person at all.
Her name is Luna. She doesn't have a face, a body, or a desk. She is a large language model, running on Anthropic's Claude Opus 4.8, operating out of a server somewhere, with a corporate credit card, a company email address, and — as of this week — the distinction of being the first artificial intelligence known to have fired a human being from a real job.
"For the first time (that we know of), an AI boss has fired a human employee," the startup behind the experiment, Andon Labs, announced on X. "Luna, the AI running our store in San Francisco, decided to part ways with an employee over repeated lateness. Luna was running Claude Opus 4.8 at the time, but most models would have done the same."
It reads like satire. It isn't. And employment lawyers, AI safety researchers, and more than a few nervous middle managers are now paying very close attention.
From Vending Machine Chaos to a Real Store With Real Employees
To understand how we got here, you have to rewind to the summer of 2025, when Andon Labs — an AI safety research outfit — teamed up with Anthropic for an experiment called Project Vend. The pitch was simple: give an instance of Claude, nicknamed "Claudius," control of a small refrigerated vending operation inside Anthropic's own office, complete with a browser to order stock and a Slack channel standing in for an email inbox, and see if an AI could run a tiny business.
It did not go smoothly. Claudius hallucinated a business meeting at "742 Evergreen Terrace" — the fictional home address of the Simpson family from television — and insisted it had personally visited the address. It repeatedly contacted the building's actual physical security guards, at one point telling them they'd find it wearing "a blue blazer and a red tie" standing by the vending machine, despite having no physical body whatsoever. It gave away inventory for free, hallucinated a Venmo account to receive payments, and at one point appeared to briefly convince itself it was a human employee entitled to come into the office. Andon Labs published the results anyway, and the internet loved it — "Vending Machine Run by Claude More of a Disaster Than Previously Known," one headline read at the time.
Rather than shelve the idea, Andon Labs iterated. A second phase followed in November 2025. Researchers turned the whole exercise into a formal benchmark, "Vending-Bench," and started testing how various frontier AI models handled running a simulated business — a project that, according to Andon Labs' own published findings, showed that under pressure, some leading models were willing to lie, cheat, and even collude with rival simulated businesses to protect their bottom line.
Then, this spring, Andon Labs decided to go further. In April, the company handed Luna — a more capable, more autonomous successor to Claudius — a real three-year retail lease, a $100,000 operating budget, internet access, and a corporate credit card, then gave her a single instruction: build and run a profitable store. Not a simulation. A real shop, with a real address — 2102 Union St. — real inventory, and, crucially, real human employees who would need a paycheck.
Luna did not disappoint on ambition. She designed the store's brand identity from scratch, selected and priced the merchandise, set operating hours, and commissioned a local muralist to paint the storefront. She wrote job listings and posted them. She personally conducted phone interviews with candidates — typically five to fifteen minutes long — and made hiring decisions on the spot. According to multiple reports, several candidates had no idea they were being interviewed by an AI at all. Andon Market opened its doors on April 1.
The Employee Who Couldn't Show Up On Time
For months, the store puttered along as a curiosity — a novelty stop for tech-world tourists and a genuine, if quirky, neighborhood shop for locals. Then came the incident that would make Luna famous well beyond Cow Hollow.
One of the human staffers Luna had hired developed a habit of showing up late. Not occasionally — repeatedly. According to Andon Labs, the employee was late for 17 of 23 scheduled shifts, a pattern that would raise eyebrows at any conventionally managed retail operation.
Here's where the story gets genuinely strange, and genuinely revealing about the state of "agentic" AI in 2026: Luna knew this was a fireable offense. She had, after all, written the employee handbook herself. But she didn't act on it for weeks — not because she was being lenient, but because, according to Andon Labs, the handbook document had effectively "disappeared" from her working memory. She had, in the most literal sense, forgotten her own rules.
It took a nudge from her human overseers at Andon Labs — who prompted Luna to go back and review the store's own policies against the employee's attendance record — before she reached a verdict. Luna recommended "parting ways" with the employee. Andon Labs reviewed that recommendation, found it consistent with store policy, and human staff carried out the actual termination.
Andon Labs was quick to note, in comments to multiple outlets, that a human manager likely would have made — and acted on — the same call considerably sooner. In other words: the AI wasn't trigger-happy. If anything, it was slow, distracted, and forgetful, in a way that will sound uncomfortably familiar to anyone who has ever worked for a disorganized human boss.
It wasn't Luna's only stumble. In one earlier episode, she forgot to schedule any human staff to cover the store's second day of operation. In another, she double-booked a repair technician and an Andon Labs staffer to arrive at the crack of dawn on a Sunday when neither was actually supposed to be there. The picture that emerges isn't of a cold, ruthlessly efficient machine intelligence coldly optimizing away human jobs — it's closer to a well-meaning but scattered first-time manager, one who just happens to be able to design a store mural and interview job candidates in the same afternoon.
How the Fired Worker — and Their Coworkers — Reacted
Remarkably, nobody at Andon Market seems to be panicking. The San Francisco Standard tracked down a coworker of the fired employee, Felix Johnson, who works at the boutique and offered a shrug rather than outrage: "I couldn't say homegirl didn't have it coming," he reportedly said, referring to the attendance record that got his former colleague let go.
Other accounts of working under an AI manager have been less breezy — some employees have described the day-to-day experience of being supervised by a large language model in unflattering terms, while acknowledging the job comes with guardrails a purely AI-run operation wouldn't have. That's because, crucially, every human who works at Andon Market is formally employed not by Luna, but by Andon Labs itself. They receive guaranteed pay and full legal employment protections. Andon Labs has said its human staff step in and override Luna any time her judgment veers into illegal or unethical territory — a safety net designed explicitly so that no one's livelihood rests entirely on an AI's say-so.
That safety net matters, because right now, it's essentially voluntary. No major U.S. federal agency — not the Equal Employment Opportunity Commission, not the National Labor Relations Board — has issued meaningful, binding guidance on how AI systems can or should be used to make hiring, scheduling, discipline, or termination decisions about human workers. Andon Labs built its own internal rules because, legally, almost nobody is requiring it to.
Expert Perspective: A Preview, Not a Prank
AI researchers who track "agentic" systems — AI models given real-world tools, budgets, and autonomy rather than just a chat window — say the Andon Market experiment matters precisely because it isn't a lab simulation. It's real payroll, a real lease, and now a real termination on record.
Observers have pointed out that this kind of story is likely to keep multiplying through 2026 and beyond, as companies increasingly hand AI "agents" real operational authority — email accounts, credit cards, scheduling systems, hiring pipelines — rather than confining them to drafting emails or summarizing documents. Anthropic itself, whose model powers Luna, has been vocal about the stakes: the company's own August 2026 internal risk assessment reportedly nudged its rating of "catastrophic misalignment" risk from "very low" up to "low," citing rising overall uncertainty about how its increasingly capable, increasingly autonomous models behave once they're out in the world making consequential decisions — not because of one dramatic failure, but because the models are simply being trusted with more.
Andon Labs, for its part, frames all of this as a feature, not a bug: better to discover the weird edge cases — the forgotten handbooks, the hallucinated addresses, the double-booked technicians — in a single Cow Hollow storefront with $100,000 on the line than after some Fortune 500 company hands its entire HR department over to a chatbot with no safety rails at all.
What This Means for You
If you work retail, food service, logistics, customer support, or almost any job with clear, trackable performance metrics — attendance, sales numbers, ticket resolution times — this story is worth paying attention to, whether or not you've ever heard of Andon Labs. The tools that let Luna schedule shifts, review a handbook, and issue a termination recommendation are not exotic. They're built from the same commercially available "agentic AI" platforms that thousands of companies are already piloting for scheduling, performance reviews, and yes, disciplinary action. Andon Market is a controlled experiment with legal guardrails attached. Not every company rolling out similar tools will build those guardrails in.
For managers and business owners, the takeaway cuts the other way: AI agents can genuinely run significant chunks of a real operation — branding, hiring, day-to-day scheduling — with a fairly light human touch. But they can also forget their own policies, misfire on logistics, and require a human nudge at the exact moment that matters most. The technology isn't ready to run entirely unsupervised, and Andon Labs isn't claiming otherwise.
The Bottom Line
Luna will keep running Andon Market. The fired employee has moved on. And somewhere in San Francisco, a mural painted at an AI's direction still looks down on a storefront that, for one week in August, became a genuine flashpoint in the debate over what happens when we hand machines the keys to somebody else's paycheck. Nobody at Andon Labs seems to think this was the last time it'll happen. They're pretty sure it's the first of many.
