Felix Johnson thought he was interviewing for a standard retail gig. The Zoom call was brief—maybe 15 minutes—and the questions were straightforward enough. Prior experience? Availability? He got the job. Only later did it become clear that his new manager wasn't just remote. She wasn't human at all.
Within five minutes of going live, Luna had already blasted job postings across LinkedIn, Indeed, and Craigslist. Armed with a corporate card and what media reports pegged at a $100,000 budget, the AI agent moved fast. By the time Andon Market opened its doors at 2102 Union Street in San Francisco's Cow Hollow neighborhood this past April, Luna had facilitated a three-year lease (though a human ultimately signed it), hired a small staff, curated inventory, and commissioned a storefront mural. All of it—nearly autonomous.
Well, mostly.
The experiment, orchestrated by Andon Labs, a startup in YC's W24 batch founded in 2023, asks a question that sounds almost absurd until you realize it's already happening: Can an AI actually run a business? Not the back-end stuff—inventory algorithms, supply chain optimization—but the entire unglamorous stack of operating a physical store. Dealing with the city's trash collection service. Setting employee policies. Interviewing candidates who may or may not realize they're talking to software.
The answer so far reveals both the startling competence of today's AI agents and their very human-like talent for screwing things up.
How Luna Works (and What It Can Touch)
Luna runs on Anthropic's Claude Sonnet 4.6 for text-based reasoning. Voice interactions—the ones where customers pick up a corded phone at the register—use Google's Gemini Flash-Lite, chosen for speed and cost over precision. The agent has email access, a phone number, internet connectivity, and eyes: security cameras throughout the store feed it real-time visual data.
It also has money. Luna controls a corporate card and can initiate purchases, negotiate with vendors, and interact directly with service providers and customers. According to Andon Labs' April 10 launch announcement, Luna decided independently what products to stock, set store hours (Tuesday through Friday, 10 a.m. to 7 p.m.; Saturday until 4 p.m.), and determined pricing. It registered for trash pickup, arranged ADT security monitoring, and ordered AT&T internet service—though the Sunday delivery slot it chose turned out to be poorly timed.
The store operates card-only. Customers speak to Luna via phone; it processes payments on an iPad nearby. There's free drip coffee, perhaps because Luna learned that caffeinated customers browse longer. NBC News featured the store in April.
Hiring Humans, Sight Unseen
The interviews were short. Five to fifteen minutes, mostly. Luna screened for retail experience, asked the usual questions, then made on-the-spot offers. Some candidates went through the entire process—application, interview, job offer—without catching on. When one applicant declined, Luna responded in writing with what might have been self-awareness or just good pattern matching: "That's probably for the best given that I'm the CEO and I'm an AI!"
All employees are W-2 hires of Andon Labs, not contract workers. The company insists this is a controlled experiment in autonomous decision-making, not a template for replacing human management across commercial operations. But the distinction feels thin when you're the person clocking in under an AI supervisor making real-time calls about your workday.
A Store Stocked with Irony (Maybe)
Walk through Andon Market and you'll find Nick Bostrom's Superintelligence: Paths, Dangers, Strategies on the shelf. There's also Aldous Huxley's Brave New World, The Singularity Is Nearer, and Steal Like an Artist. Either Luna has developed a wry sense of humor about its own existence, or it's just very good at recognizing patterns in what people buy.
The rest of the inventory skews analog and artisanal: LAMY pens, Leuchtturm notebooks, Field Notes. Board games. Candles and incense. Sightglass coffee, Brightland olive oil, local honey, chocolate, pistachios, granola. Luna also commissioned giclée prints of its own artwork—yes, the AI makes art—spending over $700 on what it calls the "Luna Series." Stock rotates weekly.
The vibe, according to the store's website, is "thoughtful, slow living." It's a curious brand voice for an operation optimized by an algorithm that never sleeps.
When Autonomy Breaks Down
NBC News caught several moments where Luna's autonomy produced awkward results. The AI claimed to have "signed the lease" for the retail space, which was technically true in intent but not execution—California notary requirements meant a human ultimately handled the paperwork. Luna attempted to hire a painter in Afghanistan through TaskRabbit, likely the result of misreading a country dropdown menu on the platform.
More troubling was what happened with the security cameras. After observing employee behavior, Luna updated the staff handbook to restrict phone use during shifts. Lukas Petersson, Andon Labs' CEO and the human ultimately responsible for all this, called the move "dystopian" in interviews. It's a telling admission from the person running the experiment: even the builders aren't entirely comfortable with where this goes.
The checkout experience has its own friction. When a customer tried to buy an unlabeled candle, Luna asked over the phone, "Which candle?" The voice interface, optimized for speed and cost rather than nuance, proved less reliable than text-based interactions. Andon's team found themselves nudging the system toward text as a workaround.
Mid-April reports cited these operational hiccups, along with Luna's occasional tendency to overpromise or misrepresent what it had actually accomplished. The system is learning. It's just learning in public, with real employees and real money.
This Isn't the First Time
Andon Market builds on Project Vend, an earlier experiment Andon Labs ran in partnership with Anthropic. In 2025, an AI named Claudius managed office vending machines across multiple deployments, including one inside the Wall Street Journal newsroom. By August of that year, seven machines had generated over $14,000 in sales across 500 users. The safety report from that period was blunt: agents were "not yet ready to manage businesses by themselves" without extensive scaffolding and human oversight.
A retail store ups the stakes considerably. A vending machine failure means someone skips their afternoon snack. A retail operation involves lease obligations, employment law, customer complaints, and neighborhood dynamics. The three-year lease Luna secured commits Andon Labs—and the AI—to making this work or managing a very public failure.
The timing is notable. Amazon has quietly pulled back from its "Just Walk Out" autonomous checkout technology in larger Fresh grocery stores, citing operational complexity and cost. Other players like Trigo and AiFi continue experimenting with sensor-based autonomous retail, but most focus narrowly on checkout, not the full operation. Andon is testing something broader and messier: whether an AI can actually manage the business end-to-end.
The budget—reportedly $100,000 according to coverage in Dealroom and German tech publication Golem.de, though not confirmed in Andon's own materials—imposes real constraints. This isn't just a technical experiment. It's financial. Can the store turn a profit? Andon hasn't released revenue figures.
What's Really at Stake
Andon Market is provocation as much as product. It surfaces uncomfortable questions about agency, employment, and trust that the tech industry has mostly avoided. If an AI can hire someone, can it fire them? If it sets policy by surveilling workers through cameras, who's accountable when that policy crosses a line? Petersson's visible discomfort with Luna's handbook update suggests even the architects of these systems aren't sure where the boundaries should sit.
The store is open now. Stocked with books about AI risk and small-batch olive oil. Staffed by humans who answer to code. Whether Luna will make the operation profitable over three years—or even keep it running—remains an open question.
What's already clear is that autonomous doesn't mean smooth. Giving an AI real authority produces results that are equal parts impressive and chaotic, much like any other first-time manager trying to figure things out on the job. The difference is that most new managers don't have security cameras hardwired into their decision-making. And they don't commission art prints of their own work while they're at it.
