AI-generated code: What has to be true for us to trust it without looking at it?

The necessity of human-readable code in an AI-driven future remains a point of contention among technology leaders. While some argue that code will become an obsolete intermediary by 2028, replaced by direct specification-to-binary generation, others maintain that manual code review is essential for safety-critical systems and performance optimization. Current AI tools often struggle with junior-level engineering mistakes, such as unnecessary memory copies, and high token costs remain a significant barrier to widespread adoption. The evolution of "harness engineering"—the development of specialized skills, plugins, and guardrails—is identified as the critical path forward. By shifting focus from writing code to refining specifications and fitness functions, engineering teams aim to improve reliability and cost-efficiency, though the human element of defining clear requirements remains the most significant challenge in the software development lifecycle.