Safety analysis focused on population-level dynamics of AI agents rather than individual agents or fixed-size groups.