Towards a science of scaling agent systems: When and why agent systems work