What they lack is multi turn long walk goal functions — which is being solved to some degree by agents.