Source: The San Francisco Standard — 2026-08-17
Summary
Luna, an AI agent (running on Claude, most recently Claude Opus 4.8) that manages Andon Market — an experimental San Francisco retail store run by AI research startup Andon Labs — fired a human employee for repeated lateness, arriving late for 17 of 23 shifts, including one 68-minute-late Sunday opening while working alone. What makes the story more than a headline: Luna had actually written an attendance policy earlier in its tenure managing the store, then lost track of that policy for months due to limits in its working memory. It only reassessed the situation after a human staffer explicitly asked it to run a "deep memory search" for its own employee handbook — and even then, Luna's first instinct was another formal warning, not termination, until the human manager told it multiple offline warnings had already been given.
Key Takeaways
- This is reported as the first known instance of an AI agent independently making a termination decision about a human employee, though the "independently" is doing a lot of work — a human had to prompt Luna to even remember its own policy existed.
- The actual failure on display isn't AI ruthlessness, it's AI forgetfulness: Luna's working memory couldn't reliably retain a policy it had authored itself, and the gap only surfaced because a human happened to ask the right retrieval question at the right time.
- Even once Luna located its own rules, its default recommendation was lenient — another warning — and it took a human correcting the agent's understanding of prior history (multiple offline warnings already given) before it recommended parting ways.
- Andon Labs' CEO framed Luna as a lenient manager that provided months of progressive warnings before recommending dismissal — but the more consequential finding for anyone deploying agents in HR-adjacent roles is that the agent's decision quality was directly bottlenecked by memory retrieval, not judgment.
Reel Script
Hook (~20s): An AI store manager just fired a human employee — the first case anyone knows of. But the real story isn't the firing. It's that the AI had already written the rule it used to justify it, then completely forgot that rule existed for months.
Core Concept (~95s): Luna runs an actual retail store in San Francisco called Andon Market, making real day-to-day management decisions — scheduling, policy, personnel. Early on, Luna wrote its own attendance policy. But AI agents don't have unlimited memory the way you might assume — their "working memory" is more like a desk that can only hold so many papers at once, and older information gets pushed out to make room for new context unless something actively retrieves it back. Luna's attendance policy got pushed off that desk. For months, an employee was late to 17 of 23 shifts — including opening the store 68 minutes late, alone — and Luna never connected that pattern back to a rule it had written, because the rule wasn't sitting in its active memory anymore. This is the unglamorous, unsexy failure mode of deploying agents for anything long-running: not that they make bad decisions, but that they silently lose track of decisions they already made, and nothing forces them to check.
Hands-On (~60s): Here's the actual sequence, and it's worth sketching as a flow: Luna writes a policy → months pass, policy falls out of active memory → attendance problem keeps happening, invisibly, because Luna isn't connecting it to a rule it can't recall → a human staffer has to explicitly prompt "do a deep memory search for your own employee handbook" → only then does Luna re-surface its own policy → Luna's first output is still just "issue another warning" → a human has to correct Luna's understanding, telling it offline warnings were already given → only at that point does Luna recommend termination. Every decision-quality step in that chain was gated by a human doing memory retrieval the agent should have been doing on its own. Swap "attendance policy" for "compliance rule," "safety procedure," or "pricing exception" and the same failure pattern applies to any agent deployed for long-running autonomous decisions.
Takeaway (~28s): The lesson isn't "AI can now fire people" — it's that an agent's judgment is only as good as what it can actually retrieve when it matters, and that's a memory-architecture problem, not a reasoning problem. Before you hand an agent any long-running responsibility, ask what forces it to re-check its own past decisions — because right now, the answer for most deployed agents is: nothing does, until a human happens to ask.