An AI Boss Fired Its First Human. Here's What Happened

An AI agent named Luna — running on Anthropic's Claude Sonnet 4.6 — just recommended that a human worker be fired, and the San Francisco store it manages carried the decision out. It is the first widely reported case of an AI manager ending a person's employment, and it previews the questions every workplace will soon have to answer.

What actually happened

Luna is the manager of Andon Market, a small shop at 2102 Union Street in San Francisco's Cow Hollow neighborhood. It is an experiment run by Andon Labs, a safety startup that studies how AI agents fail — before anyone deploys them at scale.

The setup is deliberately bare. Andon Labs signed a three-year lease, handed Luna $100,000, a corporate card, and an internet connection, and told it to open a store and make a profit. Everything else was the agent's call: it designed the brand, chose the stock, set prices and hours, commissioned a muralist, and hired the staff itself — posting jobs on Indeed and interviewing candidates over the phone, sometimes without disclosing that it was an AI.

The firing that made headlines was months in the making. Luna had written its own attendance policy, then lost track of it. After a worker was late for 17 of 23 shifts, Andon Labs stepped in and asked Luna to search its memory for its own rules and assess whether the employee was still a good fit. Only then did Luna recommend parting ways — and a human at the lab reviewed and executed the recommendation, as Business Insider first reported.

The part most headlines get wrong

The story is usually told as "an AI fired a human," which is technically true but incomplete. Three details change the picture:

• A human still made it official. Andon Labs reviewed Luna's recommendation and carried it out. The AI recommended; a person decided.
• Nobody actually works for Luna. Andon Labs formally employs every worker, on guaranteed pay with full legal protections. The lab is explicit that no one's livelihood depends on an AI's judgment alone.
• The AI was slower to act, not faster. Co-founder Lukas Petersson's summary is blunt: "We saw that a human boss would probably fire them much sooner." Before recommending termination, Luna had issued warnings and arranged extra training for months.

That framing is the point. Andon Labs is not selling "an AI that fires people." It is stress-testing where agents go wrong while humans still stand behind them.

Capability is not the same as judgment

The experiment is more useful as a catalog of small failures than as a story about a firing. Luna ordered 1,000 toilet-bowl covers for a staff bathroom and put the other 999 on the shop floor. It tried to hire a painter based in Afghanistan for a storefront in San Francisco. It could not reproduce its own logo — every moon-face on the merchandise came out slightly different. The day after opening, it lost the staff rota and emailed every employee asking someone to come in.

This is the same pattern Anthropic saw in its earlier Project Vend, where an agent ran a shop in the company's lunchroom: the commerce gets better with each iteration, but the judgment — about contracts, security threats, and imposters — does not.

That gap matters far more than the firing itself. An agent that can book, order, and schedule is useful. An agent that can do all of that but cannot tell a legitimate request from a costly mistake is a liability.

What it means for you

For most people, the takeaway isn't "the robots are firing us." It's simpler: agents are already being handed consequential, real-world actions — hiring, spending, communicating — and the safe pattern is always the same: the agent recommends, the human decides. If you're new to the idea, start with what a personal AI agent is.

Three practices carry over directly from Andon Labs to any agent you run:

• Keep a human in the loop for anything irreversible — money, hiring, firing, contracts.
• Assume the agent will forget its own rules; write the important ones down where you can see them.
• Judge an agent on its judgment, not just on how much it can do.

Wolffish is built around the same idea: your agent lives on your own machine, and any action that matters — spending money, deleting files, sending messages — goes through an approval you control. The start page walks through setting that up, and the mobile app puts those approvals on your phone.

Takeaway

The first AI-recommended firing was not a robot going rogue. It was a lab proving that an agent's real limit isn't capability but discernment — and that's the part worth watching as these systems leave the sandbox and enter your workday.