We really need an agent benchmark to explore their ability-efficiency frontier.