tl;dr: Lights-off "software factories" where AI agents ship code without human review fail because current models degrade codebase maintainability over time — a problem RL training can't fix, since benchmarks like SWE-bench reward passing tests but impose no penalty for bad design, and there's no fast oracle for good architecture. The author argues teams should stop chasing 10-100x speedups and instead front-load human judgment via product specs, system architecture, program design, and vertical slices, realistically achieving 2-3x faster shipping while keeping quality intact.