OpenAI Ships Million-Line Product Written Entirely by Codex Agents

OpenAI's engineering team spent five months building an internal beta product using exclusively Codex-generated code, producing roughly one million lines across application logic, tests, infrastructure, and documentation without a single manually-written line. The experiment required redefining the engineering role from writing code to designing environments, specifying intent, and building feedback loops that enable agents to work autonomously for hours. Key innovations included making the application UI, logs, and metrics directly legible to agents, and replacing monolithic agent manuals with a structured repository knowledge base. The team reports shipping in about one-tenth the time of traditional development, with throughput increasing as more human engineers focused on steering agent execution rather than coding directly.