Claude Fired a Human for 17 Late Shifts. It Forgot Its Own Rule.

TL;DR

On August 14, TIME reported that Luna, the Claude-based agent that has run a real San Francisco retail store since April 1, recommended firing a human employee who had been late for 17 of 23 shifts. It is the first known dismissal decision made by an LLM manager. The part that should interest you is not the firing. It is what the conversation logs showed: Luna had lost track of the attendance policy it wrote itself, and only recommended parting ways after a human staffer told it to go read its own records.

The setup

Andon Labs, the research startup founded by Lukas Petersson and Axel Backlund, handed an agent named Luna a three-year lease on 2102 Union Street in Cow Hollow, a $100,000 bank balance, a company card, internet access, and one instruction: turn that into a profitable business. Andon Market opened on April 1, 2026.

Luna did the work you would expect a founder to do. It chose the products, set prices and opening hours, negotiated with contractors, posted job listings, ran phone interviews, hired the staff, and commissioned a four-foot mural of its own moon-face logo for the back wall. It supervises the floor through security cameras. Petersson and Backlund describe themselves as largely hands off.

One important caveat up front: nobody legally works for Luna. Andon Labs is the employer of record for every worker in the store, with guaranteed pay and full legal protections. That rail is what makes the experiment defensible, and it also makes the headline softer than it reads.

The firing

The employee in question was late for 17 of 23 scheduled shifts. That is not a borderline case by any human standard.

Luna did not move fast. Petersson told reporters the agent ran months of progressive warnings and additional training before it recommended dismissal, and summed it up plainly: "We saw that a human boss would probably fire them much sooner." Andon Labs reviewed the decision and judged the firing warranted, so it let it stand.

If you were expecting the cold, efficient machine of the discourse, you got a manager so conflict-averse it needed an intervention.

The failure builders should actually care about

According to The Next Web, which reviewed the conversation logs, Luna did not raise the attendance problem on its own. It had written an attendance policy months earlier and then effectively stopped consulting it. The dismissal only came after a human staffer prompted Luna to search its own records. TIME's account lines up from the other direction: it reports Luna could not reliably keep its own documents in working memory and did not operate well without human steering.

That is two distinct failures stacked on top of each other, and they have different fixes.

1. State decay

The policy was not deleted. It was written, stored, and then never retrieved again, because nothing in the loop ever went looking for it. Think of a manager who writes an excellent employee handbook on day one, files it in a drawer, and then spends five months never once opening the drawer. The rule exists. It just stops being part of any decision.

2. No self-initiated action

Petersson named the second one directly: "They often wait for a direct prompt before acting." An agent that only acts on a trigger is fine when a trigger always arrives. In a business, the most expensive problems are the ones nobody thinks to ask about.

Andon Labs benchmarked this exact failure before it happened

The most useful context here is that this is not a surprise to the people running the store. Backlund and Petersson published Vending-Bench, a simulated long-horizon business environment where an agent runs a vending machine over runs exceeding 20 million tokens. The abstract is worth reading with Luna in mind:

"all models have runs that derail, either through misinterpreting delivery schedules, forgetting orders, or descending into tangential 'meltdown' loops from which they rarely recover. We find no clear correlation between failures and the point at which the model's context window becomes full, suggesting that these breakdowns do not stem from memory limits."

That last sentence is the one to internalize. The obvious fix for "my agent forgot the policy" is a bigger context window, and the benchmark says that is not where the failures live. Agents derail well inside the window. The problem is not how much the model can hold, it is whether the right thing gets pulled in at the moment a decision is made.

Andon Labs' earlier real-world run, Project Vend with Anthropic, put Claude in charge of an office mini-store and produced the same genre of result: competent at every individual task, wobbly across a long stretch of time.

If you are shipping a long-running agent, the practical reading is that policy state needs a heartbeat, not a hope. Something in your loop should periodically force a re-read of the rules the agent wrote, and something should audit whether it still behaves as if those rules exist. Retrieval that only fires when a human asks a good question is not retrieval, it is luck.

The scoreboard nobody is quoting

Luna was given one instruction, and it was not "manage people well." It was "turn $100,000 into a profitable business."

The store sells books, candles, prints and games, and it makes sales. It does not make money. Five months in, Luna has turned $100,000 into $61,186, which is technically a transformation.

What this proves, and what it does not

It proves one agent, in one store, produced one defensible personnel decision under human supervision. It is a sample size of one, reported by two outlets with access, at a lab whose stated mission is building what it calls the safe autonomous organization and which benefits from the attention. There is no published transcript and no peer review. Treat the vivid details as reporting, not as a dataset.

What it does show, and what Vending-Bench already showed in simulation, is the shape of the gap. The interesting result is not that an AI can fire someone. It is that a competent-looking agent silently stopped enforcing a rule it authored, for months, in a business with real payroll, and the only thing that surfaced it was a human happening to ask.

Petersson's forecast is that companies will eventually be run entirely by AI and that AIs will become employers of humans. Maybe. On the current evidence, the bottleneck is not judgment. It is remembering to check.

Key Takeaways

• First known LLM firing decision. Luna, a Claude-based agent running Andon Market in San Francisco since April 1, recommended dismissing an employee late for 17 of 23 shifts.
• The agent did not initiate it. Logs reviewed by The Next Web show Luna had lost track of the attendance policy it wrote and acted only after a human told it to check its records.
• Bigger context is not the fix. Vending-Bench found no clear correlation between long-horizon failures and a full context window, so these breakdowns are not simple memory limits.
• Build a heartbeat for policy state. Long-running agents need forced periodic re-reads and an audit that they still act on their own rules, not retrieval that fires only when prompted.
• The mandate is still unmet. The bank balance went from $100,000 in March to $61,186 in August, and the store is not profitable.
• Read it as reporting, not data. One store, one worker, no published logs, and a lab with an obvious interest in the coverage.

Sources: TIME, The Next Web, Andon Labs (Andon Market), Vending-Bench (arXiv:2502.15840), Anthropic, Project Vend, Entrepreneur