An LLM store manager recommended firing a human employee, and the reason is a context bug.

Andon Labs' agent Luna, built on Claude Opus 4.8, has run the Andon Market storefront in San Francisco for five months. After an employee missed or was late for 17 of 23 scheduled shifts, Luna recommended the company "part ways" with them, and a human team executed it (The Next Web). The detail that matters: Luna initially tolerated the lateness because the employee handbook had fallen out of its working memory, and only escalated after a human manager reminded it the policy existed. That's a context-management failure producing a materially different HR outcome. It's also the strongest concrete argument I've seen for treating agent memory retention as a compliance control rather than an engineering nicety, and it rhymes exactly with the coherence-debt paper: the missing fact didn't produce a stall, it produced different work.