Subagents are isolated context windows, which means they cannot get polluted as easily with garbage from the main thread. You can have multiple of them running in parallel doing their own separate things in service of whatever your own “brain thread”… it’s handy because one might be exploring some aspect of what you are working on while another is looking at it from a different perspective.

I think the people doing multiple brain threads at once are doing that because the damn tools are so fucking slow. Give it little while and I’m sure these things will take significantly less time to generate tokens. So much so that brand new bottlenecks will open up…