I think this is just anthropomorphism. Sub agents make sense as a context saving mechanism.
Aider did an "architect-editor" split where architect is just a "programmer" who doesn't bother about formatting the changes as diff, then a weak model converts them into diffs and they got better results with it. This is nothing like human teams though.
Absolutely agree with this. The main reason for this improving performance is simply that the context is being better controlled, not that this approach is actually going going to yield better results fundamentally.
Some people have turned context control into hallucinated anthropomorphic frameworks (Gas Town being perhaps the best example). If that's how they prefer to mentally model context control, that's fine. But it's not the anthropomorphism that's helping here.