Agent swarms have figured out how to mostly work around context window limitations. There's no reason the same ideas can't be applied on a smaller scale, at least in principle.