I bet internal use of these LLMs is insane with crazy context windows and insane compute. I was thinking about why they haven't thought about subagents already - it's been user driven because they haven't had the issues the masses have had. So it was overlooked