Drawing a parallel between compiler output and AI agent output, this piece argues that the real problem with lights-out codebases isn't the volume of AI-generated code — it's that we haven't built the verification infrastructure to make trusting that output reasonable. Just as we trust compilers not blindly but through surrounding apparatus (type systems, test suites, monitoring, rollback), we need equivalent upstream formal specifications, comprehensive AI-checks-AI pipelines, and mature production instrumentation for agent-generated code. The post frames the lights-out codebase not as something to fear or accept passively, but as a design target that reveals exactly what engineering infrastructure still needs to be built.