Adrian Quiroga discusses responsibility for AI-built systems after Codex-generated code passed functional tests but triggered a complexity check. Rather than memorizing every line or ignoring implementation altogether, he proposes fundamentals such as data structures and networking, broad knowledge to locate problems, and the ability to investigate unfamiliar code deeply. He also describes redesigning agent lifecycle handling around a state-machine architecture. This is a practitioner’s argument, not a controlled productivity study, and connects with the established idea that failures can leak through abstractions.