TL;DR: A May 2026 paper separates two capabilities that self-improving agents usually conflate: writing harness updates and benefiting from them. Writing is flat across model tiers: a 9B open model produces updates roughly as useful as a frontier model. Benefiting is an inverted-U, where mid-tier models gain most, strong models sit near the ceiling, and weak models can’t follow the harness they’re given. The practical move: put your cheap model in the evolver seat and your expensive model in the solver seat. I reproduced the mechanism on Amazon Bedrock, where a Haiku-written skill lifted a Sonnet solver from fail to pass.