Self improving AI agents are designed to learn from mistakes, but new research suggests that failure can sometimes reinforce ...