Efficient RL via Disentangled Environment and Agent Representations