Researchers propose a co-evolving framework for AI agents to learn from failures by generating hard negatives. This framework improves average task reward by 5.7% across three domains. The code will be made publicly available.