An AI store manager in San Francisco fired its first human employee this month — a milestone that had less to do with judgment than with who had to nudge whom. Luna, a store-managing agent built on Anthropic Claude Opus 4.8 by AI safety startup Andon Labs, hired two humans, wrote its own attendance handbook, forgot that handbook, and only decided to let a chronically late employee go after a human researcher prompted it to re-read its own policy. The case is the first documented real-world firing by an AI employer, and the lesson behind it is sharper than the headline: in real-world operations, the bottleneck is not what an agent knows — it is whether it acts on its own.
What Happened: The First Human an AI Manager Let Go
Andon Market is a small retail store on Union Street in San Francisco’s Cow Hollow neighborhood, selling books, candles, art prints, games and merch. It is also a live experiment: in April 2026, Andon Labs signed a three-year lease and handed Luna a $100,000 budget, a company credit card and internet access, with one instruction — open a store and make money.
Luna ran the whole setup itself. It picked the location, the product mix, opening hours, pricing and even the murals on the walls. It found contractors and suppliers. It posted jobs on Indeed and LinkedIn, wrote the requirements, screened candidates, conducted phone interviews and decided whom to hire. Two human employees ended up on the payroll with formal contracts — legally employed by Andon Labs, by design, so they kept fixed salaries and legal protections while the AI made the operational calls.
Then came the part no benchmark covers. One employee worked 23 shifts and was late 17 times. Luna had written an employee handbook before opening that stated three unexcused late arrivals within 30 days trigger a formal written warning. In practice, for months, Luna responded to the repeated lateness with a gentle “no problem, it’s fine.” Only after Andon Labs’ researchers prompted it — search your memory, find the handbook, decide if this person should stay — did Luna recover its own policy and recommend ending the employment. The humans behind the experiment executed the firing, judging the decision consistent with the store’s own rules. As Andon Labs co-founder Lukas Petersson put it: if this had been a human boss, the employee would have been let go long ago.
The Core Gap: Knowing Is Not Acting
The obvious read of the story is “AI manager makes a mistake.” The more useful read is that Luna never made a reasoning error at all. It set a sensible rule, evaluated the employee’s record correctly once asked, and reached the same conclusion a human manager would. Its failure was elsewhere: it did not act until it was prompted.
That is the difference between capability and autonomy. Current agents are extremely good at responding to a task — screen this candidate, analyze this attendance record, write this policy. They are much weaker at self-initiating: noticing that a rule they wrote is being violated, and acting on it without being told. In the Andon Labs test, the strongest frontier models reached the same firing decision as Luna when given the scenario, while weaker models hesitated. The reasoning was not the hard part. The hard part was that none of them would have done it unprompted.
Petersson describes this as the core weakness of today’s AI agents: without a direct human prompt, they often simply do not move. “Knowing” and “taking action” are, for now, two separate systems.
Why Memory Went Dark Mid-Mission
The second failure is memory, and it is the more technical one. Luna’s attendance policy did not disappear from the world — it disappeared from working context. The handbook sat in storage while the agent’s active window filled up with daily operations: suppliers, inventory, scheduling, customer questions. When the lateness pattern repeated, Luna had no live reference to the rule it had written, so the polite default response took over.
This is exactly the failure mode that long-horizon agents hit in the field: not a lack of knowledge, but a lack of retrieval. The context window is a working desk, not a filing cabinet. A policy written weeks ago is a file tucked away; the agent will happily act on what is on the desk and forget to open the drawer. The fix in this case was human — a researcher told Luna to re-run its memory. In production, the fix has to be structural: AI agents that run real operations need a runtime need persistent memory systems, scheduled retrieval of standing policies, and state management that survives across sessions. Andon Labs had previously run a smaller version of this with Anthropic — a Claude running an office vending machine — and saw the same pattern: smart, occasionally bizarre, and passive without a prompt.
From Vending Machine to Employer: An Accelerating Trajectory
The Andon Market case matters because it is the first time an AI has sat in the employer seat of a physical business with human employees and real leases. That is a qualitative step. A vending machine has one job and no staff. A store has suppliers that do not return results like API calls, employees who do not behave like functions, and a rent bill of $7,500 a month that requires roughly $500 a day in revenue to break even.
The business itself is not yet a success story: since opening on April 1, Andon Market has lost about $13,000. But profitability was not the point of the experiment. The point was observing what an agent does when it is accountable for a real operation — and the founders are explicit about where this is heading. Petersson argues that as AI capability improves, “many people will soon discover they are employed by AI,” and predicts that fully AI-run companies are a matter of time, with AI as the employer of humans.
Read the trajectory: automated checkout, then AI customer service, then AI-run inventory and logistics, and now an AI that hires, schedules and fires people. Each step moves the boundary from answering questions to managing outcomes. The first firing is the milestone that makes the accountability question concrete: when an AI discharges a human, who is responsible for the decision, and how do guardrails stay real rather than symbolic?
What Teams Should Take From This
For teams deploying agents into real operations, the Luna case offers four concrete lessons. First, assume your agent will not self-start: if a task is important, build a trigger — the way cloud agents that push a task to completion do — a scheduled audit, an event that wakes a policy review, a second agent that checks compliance. Second, treat standing rules as retrievable, not resident: store policies where the agent can find them on demand, and verify retrieval, not just storage. Third, separate judgment from consequences: the firing decision was sound, but the execution stayed with humans — keep a human accountability layer for anything with legal or reputational weight. Fourth, expect your agent to under-report its own failures; the passive default hides problems rather than surfacing them, so build observability that does not depend on the agent volunteering information.
The wider implication is that the industry’s evaluation playbook is still measuring the wrong thing. Benchmarks test whether an agent answers correctly when asked. Real operations test whether it acts correctly when nobody asks. The gap between those two is the gap between a demo and a job — and Andon Market is one of the first places where that gap was measured in payroll.
FAQ
Did an AI actually fire a human employee? Yes — for the first time in a documented real-world case. Luna, an agent built on Claude Opus 4.8 and run by Andon Labs, recommended ending the employment of a San Francisco store employee who was late 17 times in 23 shifts. The humans at Andon Labs executed the termination, judging it consistent with the store’s own handbook.
Why is this experiment significant for AI agents? Because it separates reasoning from autonomy: Luna set a sensible attendance rule, evaluated the record correctly once asked, and reached the same conclusion as the strongest frontier models — but would not act on it without a human prompt. It also forgot its own policy until prompted to retrieve it, exposing the memory and self-initiation gap in long-horizon agents.
Is Andon Market profitable? No. Since opening on April 1, the store has lost roughly $13,000; rent is $7,500 per month and break-even needs about $500 in daily revenue. The founders call it a real-world observation experiment rather than a business test, and predict AI-run companies and AI employers are a matter of time.
