The founder’s playbook making the rounds is, predictably, a vendor pitch wrapped around a four-stage lifecycle — idea, MVP, launch, scale. Skip the framing and one honest admission is worth more than the rest of it: the MVP section quietly tells founders to adopt “architecture patterns that prevent technical debt accumulation in AI-generated codebases” and “security practices for AI-generated code.”

Read that again. The same document arguing that founders who never wrote a line of code are now shipping production apps also concedes those apps carry a tax that has to be actively managed. Both things are true. The gap between shipping and operating is exactly where the work didn’t disappear — it moved.

That tracks with what I see in production AI engineering. Code generation collapses the cost of the first version to near zero. What it doesn’t collapse is the cost of the second year: the unowned abstractions, the auth path nobody reasoned through, the agentic workflow that silently retries into a bill. “Replace founder attention with agentic workflows” is a real lever, but an agent loop without an eval harness and observability isn’t leverage — it’s an outage you haven’t scheduled yet.

The playbook is right that the barrier to a first build has fallen through the floor. It’s quieter about the fact that the barrier to a reliable build hasn’t moved much at all. The skills that used to gate writing code now gate reviewing, evaluating, and bounding what the code and the agents actually do.

The HN thread splits along the obvious line — liberation versus a debt machine. My bet: the AI-native startups that survive won’t be the ones that shipped fastest, but the ones that learned to read their own generated codebase before it read them. Which muscle is your team actually building — generation, or review?