But I feel like later on the model itself will end up learning even the harness, so I wonder if there's any point in setting up a harness well on your own. I don't know if OpenAI or Anthropic built their harness well or upgraded their model well, but I ended up throwing away almost all of the harness I set up personally, and the time I spent on it became a waste of time. It seems like wasted effort, same as prompt engineering. From the point of view of each individual user. Since an already optimized system prompt is in place.