The brain has multiple collaborating models. Ex: a person with the corpus callosum severed may choose to pick up an item, but their arm will grab something else. This is a splitting of self. It seems reasonable that AIs ought to employ multiple collaborating models. It might build self.

Who is working on this?

I think this is already how a lot of closed source foundation models are built already, hence the incredible performance. A multi-agent architecture with advanced reasoning (i.e. self prompting) all happening behind the scenes. I suspect that is another reason why they don't open it up for everyone to see.