How come?

we should've learned this from the past when people thought the concept of "subhuman" was a great idea, but we never learn from the past do we

It's sort of a hack for the short-comings of current LLMs. An LLM should provide the necessary degree of "thinking" in just one shot but as it's hard to tune, we give it sub-agents that do qa or whatever. Or are you referring to something else?

I love my subagents: parallel work and small context windows for the main thread

They are great if you use them correctly