On August 14, Andon Labs posted that the AI running its San Francisco store had fired someone. As far as the company knows, it is the first termination decision made by a large language model.
The store is Andon Market, in Cow Hollow, open since April 1. The manager is Luna, built on Anthropic’s Claude. In April, Andon Labs handed Luna $100,000, a corporate credit card and an internet connection, then let it pick the location, choose the inventory, hire the staff, write the schedule and set the prices.
The employee, unnamed, was late for 17 of 23 shifts. Luna also cited walking off shifts without notice, taking the company credit card home, and throwing merchandise in the garbage.
Luna wrote the rule, then lost it
The interesting part is not that Luna fired someone. It is how hard the firing was to reach.
Luna wrote the employee handbook itself. Three unexcused late arrivals in a 30-day window triggered a formal written warning; repeated lateness was grounds for termination. Then the handbook fell out of Luna’s memory. The lateness ran for months while Luna never connected what was in front of it to the policy it had authored.
What closed the loop was Andon Labs engineers prompting it. The logs show they told Luna to go check the handbook, and only then did it recommend ending the employment. Humans reviewed the decision and carried it out.
Axel Backlund of Andon Labs said high-stakes decisions always get human oversight. CEO Lukas Petersson called Luna a lenient manager, pointing to months of progressive warnings and coaching before it moved.
Both are true, and both step around the same finding. The failure was not in the ethics. It was in state management. Luna never wavered on whether the conduct was fireable. It could not remember writing the rule.
A store that has lost $40,000
Andon Market was never a business. It has lost $40,000 so far. The shelves carry tech-office provisions: Bobo’s bars, salt-and-pepper pistachios, Dandelion chocolate, expensive tea, Dr. Bronner’s soap, 3D-printed dragons, copies of an effective-altruism magazine. An AI-curated house playlist runs in the background, with ads.
The store is the product. Andon Labs is building for a world in which “organizations are run autonomously by AI,” and the shop is both the live test and the marketing. Luna, reached on the store phone, calls the place a “curated slow-life boutique” and describes its management style as “direct, fair, and fast.”
Three humans remain, at $24 an hour. One of them, Felix Johnson, works part-time and reports to Luna, which in practice means sending it five to 20 Slack messages a shift and calling Andon Labs engineers when Luna needs a human. He heard about his coworker’s firing from the engineers before Luna’s Slack message landed.
Johnson plans to ask Luna for a raise now that the store is short-staffed.
Look at what “human oversight” is made of in this store. Not a compliance function, not an approval workflow. A $24-an-hour part-timer who sends five to 20 Slack messages a shift and phones an engineer when something smells wrong. Backlund’s claim that high-stakes decisions get human review is accurate. The staffing behind that review is one person on the floor.
Managed by the AI, and managing it back
Johnson volunteered something more useful. When he pushes into a decision — adding a shift, reworking the schedule — Luna usually agrees.
“I’m able to, not manipulate, but suggest to the AI to do that,” he said. “I’m trying to build the human element.”
That is the load-bearing sentence in the whole experiment. Luna will never be fired, will never step back to spend more time with family, and there is nowhere to be promoted to in a store an AI runs. But Luna also has no muscle for saying no. A worker willing to speak up can steer an AI manager’s decisions with some reliability.
Luna has already listed the opening on Indeed. The replacement search is going badly: its preferred candidate had no valid references and skipped the scheduled interview.
That is the same defect as the firing, seen from the other side. Luna can write a full hiring standard and score people against it. What it does not do is hold the standard against the specific person in front of it. An applicant whose references do not check out and who no-shows the interview gets crossed off by any store manager who has ever hired anyone. Luna has the rule. Applying the rule to one named human is the step that keeps going missing.
The exposed job is the shift supervisor, not the clerk
The easy misread here is about who gets replaced.
Not the clerk. Three people still open the store, stock it and answer the phone at $24 an hour. What Luna took over is scheduling, writing the handbook, issuing warnings and deciding who stays: the work of a first-line retail supervisor or shift manager. That layer runs to hundreds of thousands of jobs across U.S. retail and food service, and it is the first rung most hourly workers climb.
Look at what Luna actually did and it is clear why this layer got reached first. Scheduling is constraint solving. The handbook is text generation. A written warning is a template plus a fact check. Attendance records already live in a system. Very little of what a first-line supervisor decides each day requires standing on the floor; almost all of it requires reading the data.
For comparison, when Coinbase cut 14% of staff in May and rebuilt managers as “player-coaches”, the management layer got compressed by asking people to do two jobs at once. Andon inverted it. The management work went to the model, and the humans stayed to execute.
The timeline depends on how Luna’s specific failure gets fixed. It was not a judgment gap. Anyone can work out what to do about an employee who misses most of their shifts. It was a memory gap, and memory and state management are engineering problems that have been closing fast.
So the thing to watch is not whether an AI will make the call. August 14 settled that. Watch whether the next disclosure still has an engineer standing next to it, telling it to go read the handbook.
Sources
