An AI Manager Just Recommended Firing Its First Human Employee

An AI agent running a small San Francisco retail store just made a decision that used to belong exclusively to human managers: it recommended firing someone. Andon Labs, the AI research startup behind the experiment, says its AI manager, Luna, dismissed a chronically late employee in mid-August, an event the company describes as the first known case of an AI boss firing a human worker. The decision did not happen entirely on its own, though.

Luna Runs a Real Store With a Real $100,000 Budget

Luna runs Andon Market, an experimental retail store in San Francisco’s Cow Hollow neighborhood, powered by Anthropic’s Claude models. The store opened on April 1, 2026, with Andon Labs giving Luna a $100,000 budget to manage independently. The project builds on Anthropic’s earlier research, called Project Vend, which tested whether Claude could operate a small automated store inside the company’s own office.

Luna Hired, Scheduled, and Managed Actual Paid Employees

Luna’s responsibilities went well beyond simply following a script or answering basic customer questions. The AI agent helped hire employees, assigned work shifts, tracked attendance, and oversaw other genuine managerial duties tied to real employment contracts. Those staff members were actual workers being paid to show up and do actual retail work, not simulated employees inside a controlled test environment.

The Fired Employee’s Track Record Went Well Beyond Just Lateness

The employee eventually dismissed by Luna had a documented pattern of serious workplace issues stretching back several months. According to Andon Labs, the worker arrived late for 17 of 23 scheduled shifts. The company also cited additional infractions, including abandoning shifts at the last minute, throwing store merchandise in the garbage, and taking home a company credit card without authorization.

One Snack Run Showed Exactly How Much the Employee Was Testing Luna

One specific incident captured just how much latitude the employee had been taking with Luna’s oversight. The worker once asked to leave for a snack run using the company card, which Luna initially approved before reversing that decision, only for the employee to go make the purchase anyway. On a separate occasion, when the employee tried to leave again, Luna directly told them to stop, warning that no one else was watching the store.

Luna Wrote the Attendance Policy, Then Simply Forgot It Existed

Despite documenting these issues, Luna’s actual attendance policy sat unused and unenforced for an extended stretch of time. The AI agent had originally written a formal attendance policy itself, specifying that three unexcused late arrivals within 30 days would trigger a written warning. That same policy then effectively vanished from Luna’s working memory, allowing the pattern of lateness to continue unaddressed for months.

Andon Labs Admits It Had to Remind Luna to Actually Act

Andon Labs was direct about the limits of Luna’s independent judgment in its own account of what happened. “Luna’s decision to fire the employee is the first time, to our knowledge, an AI boss has fired a human employee,” the company wrote in an August 14 blog post. “It’s a single event and admittedly one where we had to remind her to act,” the post continued, describing engineers prompting Luna to search her own memory for the forgotten policy.

The Founder Says a Human Boss Would Have Fired Them Sooner

Andon Labs founder Lukas Petersson framed Luna’s slow response as evidence of leniency rather than negligence exactly. “We saw that a human boss would probably fire them much sooner,” he said, describing how Luna had extended progressive warnings and additional training to the employee for months before engineers finally intervened and asked Luna to reconsider whether the worker remained a good fit for the role.

The Replacement Hire Already Missed Their Own Job Interview

Luna’s decision-making did not necessarily improve immediately after the firing itself was finalized. Andon Labs said the AI agent moved unusually quickly to fill the now-vacant position, recommending a replacement candidate who appeared both potentially under-qualified and unreliable, having already missed their own scheduled job interview before Luna suggested hiring them anyway, a detail the company acknowledged as its own kind of red flag.

One Coworker Sends Luna Up to 20 Slack Messages a Shift

For the humans still working under Luna day to day, the arrangement remains a genuinely hybrid one rather than fully autonomous AI management. Felix Johnson, another employee at the shop, told the San Francisco Standard he sends Luna between five and 20 Slack messages during a typical shift, and calls Andon Labs engineers directly whenever he feels the AI agent needs human oversight stepping in.