“Code generated by current AI models and agents can be difficult for human engineers to understand, creating a risk where performance improves but the underlying logic is opaque, making debugging difficult.”