On the limits of agency in agent-based models