Ask HN: Do subagents stop making sense at some point?

If we have really powerful models (AGI) - would it still make sense to have subagents that have access to specific tools and has specific context?

Wouldn't the overlord model just do everything?