Models that build their own agent harness still lose at code and search

Six models built their own execution harnesses across 2,207 held-out tasks, against harnesses engineered by hand. The self-built ones trailed on code and search, and matched or beat the hand-built ones on writing and machine-learning experimentation.