Thanks for this paper! Still early so not quite production, but I’ve seen positive results on tasks scaling the number of “agents”. I more or less think of agents as a task focused prompt and these agents are slight variations of the task. It makes me think I’m running some kind of Monte Carlo simulation.