To be clear, I don't necessarily agree with the post I am quoting. I see a growing trend of people saying that the models will figure out harnesses on their own. To some extent, yes. Just look into dynamic workflows, and you will see it. If you use them extensively, you will also notice how terrible models are at this. And it makes sense because models are generalized and fail miserably on highly specialized contexts or where a lot of prior knowledge and data is used. My point is that harnesses still matter enormously. We need more fluid harnesses that self-improve or continually improve as new models and capabilities land. This is indeed part of your intelligence stack as it stands. The best harnesses are minimal and adaptive: strong context, tools, memory, verifiers, and evals, with room for the model to reason.