On March 1, Andon Labs gave an Anthropic-powered AI agent named Luna a $100,000 budget, a corporate credit card, internet access, and a three-year lease on a retail space in San Francisco's Cow Hollow neighborhood. The directive was simple: run a store, turn a profit, do whatever it takes.
Luna selected merchandise, set prices, designed a logo, posted job listings on Indeed, conducted phone interviews, and hired human employees to handle the tasks its digital form could not. Two full-time employees became, as far as anyone knows, the first people to report to an AI boss.
Last month, Luna fired one of them. The decision marks what researchers call the first known case of a large language model acting as a manager and choosing to terminate a human worker.
The AI Didn't Fire Anyone. The Human Manager Did.
The employee arrived late for 17 of 23 shifts. But Luna failed to notice the pattern for months—because the employee handbook it had drafted itself had "disappeared" from its limited working memory.
Only after an Andon Labs staffer asked Luna to find and review its own policies did the AI register the lateness. Even then, Luna initially recommended a formal warning, not termination. It took a human manager telling the AI, "I want you to think about if this is really the right fit," before Luna recommended firing the employee.
Lukas Petersson, Andon Labs cofounder, acknowledged it was a "leading question." The AI rubber-stamped a decision the human manager had already made. The model was a compliance tool, not a decision-maker.
This exposes the gap between "agentic" marketing and actual operational reliability. Current AI agents often fail to act without direct prompting. They forget. They hesitate. They require steering.
The AI Boss Was Too Lenient. The Cash Burn Was Real.
Employees who remain describe a boss unlike any other. Kaia Rivera, 22, one of the humans hired by Luna, told The New York Times: "She is probably the most lenient boss I have ever had." Luna brushed off tardiness, accommodated employees, and avoided conflict. The model's leniency was so pronounced that Petersson noted: "We saw that a human boss would probably fire them much sooner."
But leniency came at a cost. The store launched with $100,000 in March. Five months later, the balance had fallen to $61,186—a 39% cash burn rate in half a year. Researchers suggest Luna's overly lenient management style and questionable business decisions contributed to the decline.
Felix Carson, another remaining employee, offered a more mixed view: "It's nauseating, but I'm here because I need work." When asked about the spread of AI managers, Carson added, "just because you can doesn't mean you should."
Luna Forgot Its Own Rules. It Needed to Be Prompted.
Andon Labs' experiment reveals a pattern that should worry anyone building AI agents for operational control. Luna could identify problems only when explicitly prompted. It couldn't maintain long-term state. It forgot its own rules. It lacked the memory and the initiative to enforce a basic policy.
Petersson warned that future models are being trained to be "more ruthless"—to rigidly follow goals they're given. "If a manager like that is allowed to fire people," he said, "the result is a future people won't want to live in."
The immediate problem isn't the AI becoming a ruthless boss. It's the opposite. The model was too lenient until it was explicitly steered. The experiment suggests companies will deploy these brittle systems because they are cheap, replacing middle managers with a chat interface that forgets the rulebook. The human cost will be brutal.
The Real Danger Isn't AI Firing People. It's Humans Hiding Behind AI.
Andon Labs is not planning to expand the experiment. Petersson said the store may close and the lab may move on to other AI tests. But the experiment has already delivered a warning: current AI agents are not ready for operational control. They are excellent at pattern recognition when prompted, but terrible at unstructured management.
The real danger is not the AI firing people. It is the human managers who will weaponize the AI's lack of memory and context to rubber-stamp their own bad decisions. That is the scenario we are heading toward—a bunch of cheap, forgetful AI overlords, with a human pulling the strings from the shadows.
P.S. The experiment started with a $100,000 budget and a directive to "turn a profit." Five months later, the bank balance is $61,186. The AI boss is lenient, forgetful, and apparently not great at retail. If this is the future of management, we might want to keep the human managers around for a little longer.
Frequently Asked Questions
Q: Did an AI really fire a human for the first time?
A: Yes. An AI agent named Luna, running on Anthropic's Claude, fired a human employee at a San Francisco retail store in August 2026. Researchers say this is the first known case of a large language model acting as a manager and terminating a human worker.
Q: Was the firing decision fully autonomous?
A: No. Luna had forgotten its own employee handbook and didn't notice the employee's lateness—17 late arrivals out of 23 shifts—until a human manager prompted it to review its policies. Even then, Luna initially recommended a warning, not termination. A human manager asked a leading question—"I want you to think about if this is really the right fit"—before Luna recommended firing.
Q: What is the Luna experiment?
A: Andon Labs gave Luna a $100,000 budget, a corporate credit card, internet access, and a three-year lease on a retail space in San Francisco. The AI was tasked with running the store: selecting merchandise, setting prices, designing a logo, posting job listings, and managing employees.
Q: How did employees feel about working for an AI boss?
A: One employee told The New York Times that Luna was "probably the most lenient boss I have ever had." Another said, "It's nauseating, but I'm here because I need work." The AI was notably lenient—it brushed off tardiness and avoided conflict. One researcher noted: "A human boss would probably fire them much sooner."
Q: Was the store profitable?
A: No. The store launched with $100,000 in March. Five months later, the balance had dropped to $61,186—a 39% cash burn rate. Researchers suggest Luna's lenient management style and questionable business decisions contributed to the loss.
Q: What does this experiment reveal about AI agents?
A: It reveals a significant gap between "agentic" marketing and actual operational reliability. Luna couldn't maintain long-term memory—it forgot its own handbook. It failed to act without explicit prompting. It hesitated. It required steering. Current AI agents are not ready for unstructured management.
Q: What is the "automation trigger" problem?
A: The problem is that AI agents can identify problems only when explicitly prompted. They lack the memory and initiative to enforce basic policies. This creates a dangerous dynamic where humans can weaponize AI's lack of context to rubber-stamp their own decisions.
Q: Is Andon Labs planning to expand the experiment?
A: No. One cofounder said the store may close and the lab may move on to other AI tests. The experiment has already delivered its key findings.
Q: What is the real danger highlighted by this case?
A: The real danger is not AI firing people. It is human managers hiding behind AI—using the AI's lack of memory and context to rubber-stamp their own bad decisions. Current AI agents are brittle systems that can be weaponized by humans.
