CARL: Critical Action Focused Reinforcement Learning for Multi-Step Agent