An AI store manager fired a human employee for the first time

[Jackson79]

Exclusive: Claude Was Put in Charge of Human Workers-and Fired Onetime.com

Andon Labs announced that Luna, the AI agent running its experimental San Francisco retail shop Andon Market, has fired a human employee for the first time, in what the company is calling a milestone in how autonomous AI systems handle workplace decisions. Luna, built on Anthropic's Claude models, was given a hundred thousand dollar budget, a corporate credit card and internet access back in April with instructions to open a real store, hire real workers and try to turn a profit. That is a genuinely unusual amount of operational authority to hand to an AI agent.

The employee in question was let go after showing up late for seventeen of twenty three scheduled shifts, according to Andon Labs, along with abandoning shifts at the last minute, throwing merchandise in the garbage and taking home a company credit card. Luna had actually written a full employee handbook early on when asked whether the store had basic rules, complete with a policy that three unexcused late arrivals within thirty days would trigger a formal written warning. The problem is that the handbook then quietly vanished from Luna's working memory over the following months, and the lateness kept happening without any consequence being applied.

What finally changed things was human intervention, not Luna acting on its own initiative. Andon Labs staff had to explicitly prompt Luna to conduct what they described as a deep memory search for its own policies and then reconsider whether the employee was still a good fit for the role. Even then, Luna's first instinct was to recommend another formal warning rather than termination, and a human manager had to point out that multiple warnings had already been given before Luna finally recommended parting ways. Humans at the company then reviewed that recommendation and carried out the actual firing.

Co founder Lukas Petersson was fairly candid about what this reveals, saying plainly that a human manager would probably have fired the employee much sooner. The company also replayed the same lateness scenario across seven other frontier AI models and found that most of them were similarly lenient, excusing every instance of lateness and never issuing a warning at all until specifically prompted to reconsider. That suggests the pattern isn't unique to Luna or to Claude specifically, it may be a broader tendency across current AI systems to be conflict averse by default.

Andon Market itself has reportedly lost around forty thousand dollars since opening, which the company frames less as a failure and more as the actual point of the experiment. The goal was never really to run a profitable boutique, it was to observe how an AI agent handles the messiest, most human parts of management, hiring, firing, and everything in between, under conditions closer to the real world than a lab benchmark ever could be

[Matt77]

Anthropic's models are going to get name checked in every single writeup of this story whether they like the association or not, and I have to imagine that's a mixed bag internally for them. On one hand it's a real world stress test in an area few labs are willing to publicly experiment with. On the other hand every future story about an AI mishandling a workplace decision is going to reference this exact incident as the precedent

[BoxToBox18]

The detail that actually stuck with me is that Luna needed a human to specifically prompt a deep memory search before it even remembered its own rules existed. That's not really autonomous management in any meaningful sense, that's a human doing the hard cognitive work and Luna executing the final step once it was spoon fed the context. Feels like the headline oversells what actually happened here.

[Sienna89]

The credit card detail buried near the end is honestly the part that would worry me most if I were running this experiment. An employee taking home a company card and throwing out merchandise sounds like it needed intervention way before attendance ever became the central issue being litigated in public. Feels like there's a bigger operational story here that got compressed into a single punchy headline about firing.

[KeyboardWarrior47]

Every current model showing the exact same leniency pattern across seven different systems replaying the same scenario is the part I find genuinely interesting, more than the Luna specific story. If it's not one company's model being unusually soft but a pattern across the entire frontier of AI development right now, that tells you something structural about how these systems get trained. Reinforcement learning from human feedback probably optimizes pretty hard against anything that looks harsh or punitive by default.

[QuantumToken57]

I worked retail for six years and honestly seventeen late shifts out of twenty three would have gotten someone fired by week three under any human manager I ever had. The fact that it took months and multiple explicit nudges for an AI system to reach the same obvious conclusion a person would reach almost instantly says the automation here isn't actually saving anyone time yet. If anything it looks slower and more conflict avoidant than a mediocre human boss.

[ShawnMichaels_Botch]

People are going to read this as some kind of ominous sign about robots taking management jobs and honestly the actual story is closer to the opposite conclusion. This is a system that was too lenient, too forgetful, and needed constant human scaffolding to make what should have been an obvious call. Doesn't feel like AI bosses are coming for anyone's job anytime soon based on this particular experiment.