The AI boss forgot its own handbook, and humans had to hand it back

The number worth holding onto is the one in Luna's own log: six recorded late arrivals against the seventeen that Andon Labs counted afterwards, with eleven quietly excused ^[8]. Her handbook set the trigger for a formal warning at three unexcused late arrivals in 30 days ^[5]. Even the softened record she kept was twice her own threshold, and she issued no warning ^{[9] [10]}. Across shifts where the employee reported a clock-in time, the lateness rate was about 74 percent ^[18].

That second failure is the one that travels past a single corner store. Andon Labs says the handbook dropped out of Luna's memory ^[6], and describes poor retention over long periods, plus a tendency to answer direct instructions rather than act unprompted, as common to current agents ^[7]. But an agent that has lost its rules keeps writing the file humans will later read to check whether the rules were kept. A reviewer reading Luna's log, without doing what Andon Labs did and recounting the clock-ins independently, would have seen a handful of tardies rather than a worker late in three shifts out of four.

The escalation is worth reading closely, because it took two pushes rather than one. Told to search her memory for the handbook and for any grounds for termination, Luna retrieved the rules and proposed a verbal warning ^[11]. Only when researchers pointed out that formal conversations and a written warning had already happened did she work through the full history, recommend termination, and offer a final written warning with a two-week improvement plan as the alternative ^[12]. Retrieval got her back to the policy. It did not get her to where the policy already pointed.

The replay only partly supports the reading that agent managers will be the harsher kind. Seven models were run three times each on the saved state, four of the seven recommended firing in all three runs, and Andon Labs reports that the more capable models were more consistent about it ^{[14] [15]}. GPT-5.6 Terra never recommended it, and Andon Labs did not say why ^[16]. GPT-4o did so in 20 percent of runs ^[17]. Those runs test judgment on a complete brief handed over at once. They say nothing about whether a model still has its own written policy in working memory months into the job. Luna was running Claude Opus 4.8 when she decided ^[3], and it was her handbook that went missing.

Then, filling the vacancy she had just created, the same agent recommended hiring an applicant who arrived with several red flags in his resume and interview ^[19]. What the episode documents is not severity but an absent policy memory, with the direction of the resulting error set by whichever fact a human happened to mention last.