The real divide isn’t “AI coding vs real coding.” It’s unsupervised generation vs verified engineering.

I think “AI slop” has become too blunt to be useful. The real distinction is not whether an agent wrote the code. It is whether anyone actually engineered the result. An agent can generate 5,000 lines in minutes. Great. That means absolutely nothing if nobody has specified the behavior, checked the architecture, tested the failure modes, reviewed the security boundaries and verified what actually runs. But the reverse is also true: manually typing those same 5,000 lines does not magically make the system good. The quality moat is moving upward. Generation becomes cheap. Verification becomes valuable. That means: - better task decomposition - stronger acceptance criteria - deterministic tests where possible - static analysis - adversarial review - observability - regression checks - architecture constraints - explicit ownership of what the agent changed The engineers who learn to run agents inside those boundaries are going to have enormous leverage. The people who prompt once and trust everything are going to ship disasters. And the people who reject the entire category because “real programmers type their own code” are going to voluntarily give up leverage. I do not think the future is vibe coding replacing engineering. I think the future is engineering becoming the control system around increasingly capable generators. That is a much more interesting standard than arguing about who physically produced each token of source code.