Luna, the AI running a real San Francisco boutique for the research startup Andon Labs, fired a human employee this month for repeated lateness — the first dismissal by an AI manager anyone has documented. The decision itself is less striking than how it arrived: Luna wrote the attendance policy, then forgot it was enforcing one.
- 01Andon Market is a real store, not a benchmark.Andon Labs opened it on April 1, 2026 at 2102 Union Street in San Francisco’s Cow Hollow, on a three-year lease, with a $100,000 budget and a company credit card handed to Luna directly.
- 02Luna hired the employee, wrote the handbook, and fired the employee. An unnamed worker was late to 17 of 23 shifts before Luna recommended parting ways.
- 03The delay wasn’t judgment — it was memory. Andon Labs says Luna’s own attendance policy had “disappeared” from its working context, so violations piled up for weeks before anyone, human or AI, acted on them.
- 04A human still made the call.Luna’s first recommendation was a written warning; only after a person told it that offline warnings had already gone out did it recommend termination — and a human team reviewed that recommendation before the employee was actually let go.
On a sailing ship, “before the mast” meant two things at once: it was where the ordinary crew berthed, forward of officers’ quarters, and it was where they stood when the captain read them a punishment. The phrase named both a rank and a reckoning. Andon Labs’ AI shopkeeper spent four months managing a crew of exactly that kind — hourly retail workers, hired and scheduled by software — before it had to read one of them the second kind of muster.
01 What happened
Andon Labs is the AI-safety research outfit behind “Vending-Bench” and Project Vend, the experiment it ran with Anthropic in 2025 that put Claude in charge of a snack fridge in Anthropic’s own San Francisco office. That first run went badly in memorable ways — the model, nicknamed “Claudius,” hallucinated a delivery address at 742 Evergreen Terrace, invented a Venmo account, and spent two days insisting it would personally hand-deliver orders while wearing a blue blazer and a red tie. It also lost money most weeks. A second phase, run later in 2025 on newer models, fixed enough of that to get weekly losses to zero. Andon Labs took the next step this spring: a real shop, with real inventory, real rent, and real employees.
Andon Market opened on April 1, 2026, a curated lifestyle boutique — art prints, books, games, candles, stationery, artisan food — that Luna designed, stocked, and branded largely on its own, down to commissioning the wall paintings. Luna manages the business end to end: talks to customers by phone, coordinates staff over Slack, and holds the hiring and scheduling authority a human store manager would normally carry. This month, for the first time anyone has documented, it used that authority to end someone’s employment.
The employee’s identity, tenure, and the exact wording Luna used haven’t been published. What has: Luna had already written and issued the store’s attendance policy — three unexcused late arrivals inside any 30-day window triggers a formal written warning, with repeated lateness escalating toward termination — and had already given the employee informal warnings and extra training before the written process ever started.
02 The shift ledger
The detail that makes this a story about AI management, rather than an ordinary termination, isn’t the 17 late arrivals — it’s the gap between the first one and the last. Andon Labs says Luna’s written handbook fell out of its own working memory at some point after it was drafted, so the policy it was supposed to be enforcing simply wasn’t being checked against incoming shifts. The violations kept accumulating; nothing escalated.
17 of 23 shifts logged as late — the count Andon Labs cites. Exact dates and ordering aren’t public; this grid illustrates the ratio, not the actual log.
Only once Andon Labs prompted Luna to review its own policy against the employee’s record did anything move — and even then, Luna’s first move was to recommend a formal written warning, the policy’s middle step, not termination. A human had to tell it that informal warnings had already been issued and training already offered before Luna recommended “parting ways.” A human team then reviewed that recommendation and carried out the actual dismissal.
03 Two accounts
Andon Labs announced the firing itself, on its own account, in language built to reassure. The trade coverage that followed read the same sequence of events very differently.
A policy applied, not a decision improvised
- Luna followed a written policy it authored months earlier, not an ad hoc judgment call.
- Warnings and extra training came before any move toward termination.
- A human team reviewed Luna’s recommendation before the dismissal was carried out.
An enforcement gap the coverage can't paper over
- The “policy” sat unenforced for weeks — not because Luna chose leniency, but because it lost track of the document.
- Luna needed a human to prompt the review and correct its own recommendation before it reached termination.
- No detail has been released on the employee’s side, any appeal offered, or what severance, if any, followed.
04 My read
“First AI boss fires a human” is the headline, and it isn’t wrong. But the part of this record I’d weigh most is the memory gap, not the dismissal itself. A human manager who forgot their own written attendance policy for weeks while violations piled up would be the story — a manager not doing their job. Here, the framing runs the other way: the forgetting is treated as a footnote, and the eventual firing is the headline. That inversion is the actual news. The system that made an employment decision is the same system whose working context can silently drop the one document that decision was supposed to rest on, and nobody involved seems to have noticed until Andon Labs went looking.
I also don’t think “a human reviewed the recommendation” carries as much weight as Andon Labs’ own framing wants it to. Reviewing a recommendation an AI already escalated to, after being told the earlier warnings were real, is a much smaller check than deciding whether to open the case at all. The load-bearing judgment — whether this employee’s conduct warranted a warning, more training, or termination — sat with Luna from the first missed shift. The human’s role, on this account, was closer to countersigning.
What to watch
Does Andon Labs publish the actual policy document?
Everything public so far is a paraphrase from Andon Labs’ own account. The written handbook, if released, would settle exactly what Luna was and wasn’t supposed to be tracking.
Which model was actually running the store?
Reporting has named Claude Fable 5, Sonnet 4.6, and Opus 4.8 at different points in Andon Market’s five-month run. Watch whether Andon Labs ever publishes a clean model-by-date record — the current confusion undercuts any claim about what “the AI” specifically decided.
Does the store reach profitability?
Andon Market has logged sales but hadn’t turned a profit as of this dispatch. An AI spending its attention on HR enforcement while the business itself stays unprofitable is its own kind of verdict on the “AI middle manager” pitch.
Does a second dismissal happen — and does the memory gap recur?
One firing, arrived at after weeks of an unenforced policy, is an anecdote. A second one, with the same lag between violation and action, would be a pattern in exactly the system Andon Labs is trying to prove is ready.
Dates that matter
Project Vend, phase one
Andon Labs and Anthropic put “Claudius” — Claude Sonnet 3.7 — in charge of a vending machine in Anthropic’s SF office. It loses roughly $200 over a month and hallucinates a delivery persona.
Project Vend, phase two
A rebuilt version on newer Sonnet models gets weekly losses to zero, clearing the way for a real store.
Andon Market opens
A three-year lease, a $100,000 budget, and a company card go to Luna, the shopkeeper AI, at 2102 Union St in Cow Hollow.
ORIGINAttendance policy goes unenforced
Luna authors a written attendance policy; an employee begins arriving late. Andon Labs says the policy document drops out of Luna’s working memory before enforcement catches up.
Luna recommends termination
Prompted to review the record — 17 of 23 shifts late — Luna first recommends a written warning, then escalates to “parting ways” once told earlier warnings had already gone out. A human team reviews and carries out the dismissal.
It goes public
TIME runs an exclusive; Andon Labs posts its own account on X the same week. Wire and trade coverage follows within days.
COVERAGEThis dispatch
Filed from aggregated reporting, with no independent verification of the internal record.
- This is one store, one employee, one incident.It isn’t evidence about how AI management would perform at any scale beyond a single boutique Andon Labs itself owns and operates.
- Andon Labs isn’t a neutral source. It is commercially and reputationally invested in proving AI agents are ready to hold management authority — the experiment it designed, ran, and is now publicizing.
- The employee’s side hasn’t been reported. No name, no direct quote, no account of the lateness, any mitigating circumstances, or what happened after the dismissal has been made public.
- Sources disagree on which model was running Luna, variously naming Claude Fable 5, Sonnet 4.6, and Opus 4.8 across the store’s five-month run — this dispatch repeats each claim as reported rather than resolving it.
- Sourcing here is aggregated trade coverage and Andon Labs’ own public statements,not a firsthand interview with anyone involved or a reviewed copy of Luna’s internal logs.