Reading generated code gets you correctness. It doesn't get you fit, and fit runs on context the agent never had. What imposed and inherited defaults cost my team, where the research lands, and why your worst code is now your highest-leverage cleanup target.

Making software portable is only half the job. What a failed deployment into a locked-down customer environment taught me about testing real capabilities, owning outcomes across boundaries, and the gap between running somewhere and running where the customer actually is.

The first engineering team isn't a smaller version of the one the company will eventually need. Why I hire for range over resemblance, give every early hire a real mandate, and build a team that can make the next team legible.

Turning a research prototype into a production system isn't a code-cleanup project. It's the work of designing a boundary between a research loop optimized for learning and a service other people can depend on.