The headline wrote itself: an AI boss has fired a human worker, apparently for the first time. It comes from Andon Labs, a research outfit running an actual corner shop in San Francisco — Andon Market, on Union Street — managed by an AI agent called Luna. Andon Labs itself put it plainly on 15 August: "For the first time (that we know of), an AI boss has fired a human employee."

It's a real milestone, and it's worth understanding — because the thing that actually happened is both less alarming and more interesting than "the machines are handing out P45s now."

What Luna actually did

Start with what it didn't do: Luna did not wake up one morning and fire anyone. The sequence, from Andon Labs' own conversation logs, goes like this.

Months ago, Luna — the Claude-based agent running the store, given a $100,000 budget, a corporate card and an instruction to turn a profit — created an attendance policy of its own. Then it lost track of it. An employee kept turning up late — 17 of 23 shifts, in the end — and Luna, having mislaid its own rule, mostly let it slide. It issued warnings. It arranged extra training. For months, it took no contractual action at all.

The turn came when humans stepped in. Andon Labs prompted Luna to search its memory for the attendance policy it had written, and to assess whether the worker was still a good fit. Even then it did not reach for the sack: Luna's first instinct was to issue another warning, and only after a human pointed out that warnings had already been given did it finally recommend "parting ways." And even that recommendation was reviewed by humans at the lab, who carried out the dismissal themselves.

So the accurate version is: an AI manager, reminded by humans of a rule it had forgotten, recommended a firing that humans then executed. That is a meaningfully different sentence from the headline.

The reveal is the leniency, not the firing

Here's the part that actually tells you something about where these systems are. Andon Labs co-founder Lukas Petersson noted that "a human boss would probably fire them much sooner." The AI wasn't a ruthless optimiser hunting for someone to cut. It was, if anything, a soft manager — forgetful, slow to escalate, more inclined to issue another warning than to act. Its defining failure as a boss was not cruelty; it was losing its own paperwork.

That fits what we know about current AI agents on long-horizon tasks. They are capable in the moment and unreliable across time — a policy written in week one is functionally gone by week ten unless something forces a lookup. Luna's headline moment was really a memory failure with a human-supplied correction bolted on.

The guardrails were the whole point

It also matters, a great deal, that nobody's livelihood was actually riding on an AI's judgment. Andon Labs formally employs all the store's staff, with guaranteed pay and full legal protections. Petersson said the lab would step in over any decision that was illegal or unethical, and simply judged that this one wasn't. Luna's authority, in other words, ended exactly where a consequential decision began. It could recommend; humans disposed.

That is the correct design for this stage of the technology, and it is the opposite of the autonomous-AI-boss image the headline conjures. This was a controlled experiment with a human hand on every irreversible lever — a demonstration of what an AI agent can draft, not what it is trusted to do.

What it does and doesn't tell us

None of this makes the milestone meaningless. An AI agent ran a real shop, managed real people, wrote real policies and reached a real personnel recommendation that humans found reasonable enough to act on. A year ago that sentence would have sounded like science fiction. The direction of travel is real, and the "first that we know of" is probably true.

But the useful takeaway is not "AI managers are here." It's a snapshot of what AI managers are like right now: capable of the mechanics, prone to forgetting their own rules, gentler than a human would be, and — crucially — kept firmly on a leash by the people running the experiment. The interesting question Andon Labs is really probing isn't whether an AI can recommend firing someone. It's whether, as these agents get more reliable across time, we will keep that human hand on the lever — or quietly start taking it off.