Clearly it can't be agents "all the way down" (in any case, the component parts of the neurons do not have agency), but the idea that it will be easier to evolv

Clearly it can't be agents "all the way down" (in any case, the component parts of the neurons do not have agency), but the idea that it will be easier to evolve a high-level agent if the components themselves are agent-like seems coherent. It would be interesting if this could be sharpened into a testable prediction, in some kind of toy model. I could see it go the other way too--for example large companies and states struggle to act as coherent agents, because the individuals and business units inside them scheme against each other, pursuing their own agendas. In order to maintain high-level agency they need to suppress low-level agency, e.g. armies training their soldiers to blindly follow orders.