A new framework called SERL enhances reinforcement learning for multi-turn agents by effectively utilizing environmental feedback to improve task success rates. This approach outperforms existing methods, achieving notable success in complex environments like ALFWorld and WebShop.