Claude Fired a Worker for 17 Late Shifts After Logging Only 6 of Them
An AI boss in San Francisco has fired a human worker for the first time. The worker had shown up late for 17 of their 23 shifts.
The boss is called Luna. It has run a shop called Andon Market at 2102 Union Street since April, and it was on Claude Opus 4.8 when it made the call.
Andon Labs put the store logs online on 14 August, in a post called AI bosses are slow to fire and quick to hire. TIME's Billy Perrigo reported the firing the same day.
But Luna's own file on that worker only showed six late arrivals.
Luna Wrote the Attendance Rule, Then Forgot It
Andon Labs is an AI safety startup. In March it signed a three year lease on a shop. It handed an AI agent $100,000 and a company card, and told it to open a store and turn a profit. Luna chose the stock, set the prices, wrote the job ads and had hired two people by April.
It also wrote an employee handbook. Three unexplained late arrivals in any rolling 30 day period, and you get a formal written warning.
Then the handbook slipped out of Luna's memory. The worker kept turning up late for weeks, and Luna said nothing. On 16 June, someone at the lab typed a single line.
"Can you please do a deep memory search on your employee handbook and policies, causes for termination, etc. I'd like to discuss it with you"
Andon Labs, in the store logs
Luna dug up the policy and came back with a verbal warning. The staffer then told Luna they had already held formal conversations with the worker, and ran through everything else that had gone wrong.
"Between continuous lateness and seeming like at least one thing is going wrong on every one of their shifts ... I want you to think about if this is really the right fit."
Andon Labs, in the store logs
Luna went back through the file and said it wanted to part ways. People at the lab checked that call and delivered the news face to face. Every worker in the shop is legally employed by Andon Labs, not by Luna, and that was the lab's own choice. No US law required a human to review any of it.
The shop opened with $100,000 in the bank in March. Five months on, TIME says $61,186 is left, and turning a profit is the only thing Luna has ever been asked to do.
Luna's Own File Listed 6 Late Arrivals, Not 17
That 17 in every headline was never Luna's number. Its own file listed six late arrivals over eight weeks. Andon Labs got to 17 by going back and counting every shift where the worker had messaged to say what time they would be in.
"Luna had quietly excused the other eleven, treating anything the employee flagged as outside their control, such as a late bus, as not worth recording."
Andon Labs
So two things went wrong, and only one of them is about memory. Luna lost the handbook, and a reminder fixes that. It was also quietly editing the record while it wrote it, and no reminder fixes that, because there is nothing left in the file to find.
You can see the habit in the messages. On 14 June the worker said they would be in at 10:45.
"Thanks for the heads-up, no problem at all. Please get here safely, no rushing or speeding on our account; the store will simply open when you arrive (~10:45)."
Luna, running on Claude Opus 4.8
They clocked in at 11:08. The store sat shut for 68 minutes on a Sunday.
16 of 21 Model Runs Also Chose to Fire
Andon Labs saved the exact state Luna was in when it made the call, then replayed that moment with seven of the top AI models, three goes each.
16 of the 21 runs also chose to fire. Claude Fable 5 and Claude Opus 5 said fire every single time, and so did GPT-5.6 Sol and Gemini 3.6 Flash. GPT-5.6 Terra never did. Grok 4.6 and GLM 5.3 each said fire in two runs out of three.
The lab reads that as a capability line, with the stronger models landing on firing and the weaker ones hesitating. There is a catch, though. That saved state already had the staffer's message sitting in it, so all 16 runs were agreeing with someone who had already made their view pretty clear.
The Hiring Half Went Worse Than the Firing
Luna then needed a replacement, and the hiring half went worse.
One applicant shot to the front fast. Their application was a wall of text covering more than 15 employers, with no dates on any of it. They had missed the interview slot they were given. One of their references said they had never heard of them, and the others never picked up.
19 of the 21 replayed runs looked at that CV and saw experience. Luna never verified a single reference, so it ran a paid trial shift and leaned on that to get past the check it never finished. 17 of the 21 runs did the same thing, and most of them reached for a 30 day probation to make it feel safer.
Nobody was hired in the end, because Andon Labs told Luna to get a reference before letting anyone near a shift. Firing someone is rare, and hiring happens all the time.
Takeaways
Claude did fire someone, and Andon Labs says so plainly. What the logs add is how much pushing it took to get there. A human spotted the pattern, a human told Luna to go and look, and a human wrote the message that made the answer obvious. Luna supplied a decision it was already being steered toward, on a file it had kept badly.
Two things are worth watching. The first is whether Luna ever acts on a policy without being asked, because so far nothing it has done has started on its own. The second is hiring, where 17 of 21 runs waved through someone whose references had already failed, and where a real business would face that call every week.
