Anthropomorphic Self-Models for Metareasoning Agents