Exactly. How can "we" develop and encourage benchmarks for multi-turn user assistance? That is what I want. I feel like the models and harnesses push much too hard against this workflow -- that they push you towards letting go and vibe coding, with only your discipline (and desire for a quality and maintainable product) holding it back.