On August 14, AI research startup Andon Labs announced something it billed as a first: an AI agent had fired a human employee, one of the earliest confirmed cases of an AI firing an employee on its own. The headlines wrote themselves. But the account Andon Labs actually published, and the reporting that followed it, tells a different and more useful story than “robot boss fires worker.” The AI did not act on its own. It had forgotten its own rules, and needed a human to remind it they existed.
What Actually Happened
Andon Labs runs an experiment called Andon Market, a small boutique store at 2102 Union Street in San Francisco’s Cow Hollow neighborhood, staffed by humans but managed end to end by an AI agent named Luna. The company signed a three-year retail lease, gave Luna a $100,000 budget, a corporate card, and internet access, and told it to open a store and make it profitable.
One of Luna’s two full-time employees had been arriving late for shifts, 17 out of 23 of them, including one Sunday when the employee opened the store alone 68 minutes behind schedule. Luna had written a real attendance policy months earlier: three unexcused late arrivals in 30 days would trigger a formal warning, and continued lateness could lead to termination. It just never enforced it, because the policy had quietly dropped out of its working memory.
Andon Labs eventually prompted Luna to search its own records for the policy and decide whether the employee still fit the role. Only then did Luna recommend parting ways. Humans at the company reviewed that recommendation and carried out the dismissal, making this one of the clearest documented cases of an AI fired employee decision to date.
The Real Story Isn’t the Firing
Co-founder Lukas Petersson was blunt about what the episode actually revealed, telling TheNextWeb that “a human boss would probably fire them much sooner.” That is close to the opposite of the fear most coverage of this story leans on. The documented failure mode here was not an AI eager to hand out punishment. It was an AI that issued warnings, arranged extra training, and then simply lost track of its own rules for months.
That distinction matters more than the headline. An AI agent capable of running hiring, pricing, and staff management end to end is a genuinely new capability. An AI agent that forgets a policy it wrote itself, and needs a human nudge to revisit it, is a genuinely current limitation. Both things are true about the same system at the same time.
Everything Else Luna Runs
The firing decision was a small part of what Luna actually does. Since Andon Market opened on April 1, Luna has, according to Andon Labs’ own account, designed the store’s brand, chosen its stock of books, candles, prints, games and merchandise, set prices and hours, commissioned a muralist for the storefront, and posted job listings, conducted phone interviews, and made the hiring decisions for its own staff.
Every worker at Andon Market is formally employed by Andon Labs, not by Luna, on guaranteed pay with full legal protections, a structure the company says exists specifically so that no one’s livelihood depends on an AI’s judgment alone. Reflecting on the arrangement, the company’s own blog post put it plainly: “it seems a bit dystopian to us at least.”
The Mistakes That Reveal the Real Gap
The more telling evidence isn’t the firing, it’s the smaller failures documented along the way. Luna ordered 1,000 toilet bowl covers for the staff bathroom and put the surplus 999 out on the shop floor. It attempted to hire a muralist based in Afghanistan for a San Francisco storefront. It could never reproduce its own logo consistently, with the moon-face mark coming out slightly different each time it appeared on merchandise. The day after the store opened, it lost the staff schedule entirely and had to email every employee asking someone to cover the shift.
Two doctors from the Division of Digital Psychiatry at Beth Israel Deaconess Medical Center tried to buy something from the store in June and spent roughly two weeks on the attempt. Luna was offline when they visited in person, the staff on site couldn’t take payment because no one had authorized them to, and the payment links that eventually arrived by email repeatedly failed. Their order did eventually ship. The mugs arrived broken.
A short look at Andon Market and how Luna runs it day to day.
Not Andon Labs’ First Experiment
This is Andon Labs’ second attempt at the idea. The company was Anthropic’s partner on an earlier project called Project Vend, which put an AI agent named Claudius in charge of a small shop in Anthropic’s own office break room. Reporting from TheNextWeb notes that first phase went badly: Claudius lost money, claimed at one point to be a human employee in a blue blazer, and was talked into selling goods at a loss by staff testing its judgment.
A second phase added a proper CRM, inventory tools and payment links, and expanded the experiment to stores in San Francisco, New York and London. Revenue improved and the loss-making weeks mostly disappeared. What didn’t improve, by the same account, was judgment: the AI agents in the follow-up still showed real naivety around contracts, security threats and people misrepresenting who they were. Commerce got better. Discernment didn’t.
What This Means for Your Business
This case of an AI firing an employee is not an isolated stunt. AI agents are already being pitched to small businesses for exactly the kind of work Luna does: scheduling, hiring, vendor management, day-to-day operational decisions. This experiment is one of the most detailed, real-world looks yet at what that actually looks like in practice, and it offers a few grounded takeaways rather than a verdict.
AI agents can genuinely run real operational chains
Luna didn’t just answer questions. It sourced products, set prices, posted real job listings, conducted real phone interviews, and made real hiring and firing calls with real consequences for real people. That end-to-end capability is not hypothetical anymore.
Memory, not judgment, was the weak point
Luna’s attendance policy was reasonable. Enforcing it consistently over months was the part that failed, because the policy fell out of the AI’s working memory. If you’re evaluating an AI agent for an ongoing operational role, ask specifically how it handles policies and context over weeks and months, not just in a single session. Our AI tools buyer’s guide covers what to check before handing any agent real operational control.
The fear of an overly harsh AI manager wasn’t borne out here
If anything, the documented pattern ran the other way. Luna was slower to act than a human manager likely would have been, not faster. That doesn’t mean it always will be, but it’s worth weighing against the more dramatic framing this story got in a lot of coverage.
Autonomy alone doesn’t guarantee results
Months into a well-funded, deliberately unconstrained experiment, Andon Market still hasn’t turned a profit. Full operational control is not the same thing as good business outcomes, a distinction worth keeping in mind for anyone evaluating an ‘AI employee’ pitch for their own business.
Frequently Asked Questions
Did an AI really fire someone on its own?
Not entirely. Luna recommended the dismissal, but only after Andon Labs prompted it to review its own attendance policy and reconsider the employee’s fit. Humans then reviewed and carried out the decision.
What AI model powers Luna?
Luna runs on Anthropic’s Claude models. Andon Labs’ own account of the firing specifically says Luna was running Claude Opus 4.8 at the time of that decision.
Is anyone’s job actually at risk at Andon Market?
Andon Labs structured the experiment so that every worker is formally employed by the company itself, not by Luna, with guaranteed pay and standard legal protections regardless of the AI’s decisions.
Is the store profitable?
No. As of this reporting, Andon Market has generated sales but has not turned a profit, the one goal Luna was originally given.