Comment by arglebarnacle

The models are getting better at agentic coding, so over time using complicated harnesses and precise prompt engineering to attempt to squeeze out an extra X% performance will become irrelevant as the models approach expert-level performance. The bitter lesson in miniature.