A team at OpenAI spent five months building a real software product using only Codex-generated code — not a single line was written by a human hand. They shipped a million lines of code with a small team in about one-tenth the time it would normally take. The product has daily users, external testers, and gets fixed when it breaks, all with agents doing the coding. The key shift? Humans stopped being coders and became system designers: they created environments, specified intent, and built feedback loops for the agents. Early on, progress was slow because the environment was too vague — the agents needed clear tools and structures to work efficiently. Once the scaffolding was right, throughput soared to 3.5 pull requests per engineer per day. Why it matters: This experiment points to a future where the engineer’s scarcest resource isn’t code output but human attention — and the bottleneck shifts from writing software to designing the systems that write it.