An AI manager fired someone after forgetting its own rule for months!
Andon Labs runs a shop in San Francisco where an AI agent called Luna does the managing: it picked the stock, set the prices, ran the hiring. This month it recommended dismissing a worker who had turned up late for 17 of 23 shifts. The logs show Luna wrote that attendance policy itself, months earlier, then lost track of it and only acted after a human at the lab intervened and asked it to go and look up its own rules. Humans reviewed the recommendation and carried it out.
What it actually means
The headline is "AI fires human". The lesson is the opposite of ruthless. Luna didn't over-reach or under-reach, but arguably you could say it sort of drifted? For months, on a rule it had written itself, until a person prompted it; imo intervened. If you are handing an agent a policy, be clear about what you have actually built: something that can apply a rule when asked, not something that notices when the rule is being broken. Enforcement still needs a trigger, like a scheduled check, an alert, a human who reads the exceptions, because the agent will not go looking. The failure mode that costs you isn't the agent doing something but humans cannot get too complacent.
An agent will apply your rule when you ask it to. It won't notice when the rule is being broken.
17 of 23shifts missed before the AI acted at all
Source: The Next Web · Wednesday, 19 August 2026