The paper proposes an empowerment-guided multi-agent system that uses semantic checkpoints and structured communication to ensure that complex scientific computing workflows maintain semantic consistency and reliable decision-making across multiple agents.