An internal experiment at OpenAI has been revealed: A team of just three engineers initially developed a product with millions of lines of code from scratch using Codex agents over five months, without humans writing a single line of manual code. In this experiment, human engineers focused on clarifying requirements and setting rules, while the rest of the development work was completed by AI, including code generation, testing, documentation writing, and CI configuration. OpenAI named this workflow 'Harnessing Engineering.' The experiment showed that AI is most efficient in environments with clear boundaries and predictable structures, where humans interact with the system through prompts to advance the completion of code merge requests (PRs). As the code volume increased, the bottleneck in AI coding shifted to human quality inspection capabilities, which OpenAI overcame by enabling Codex to directly read the application's user interface, logs, and application metrics.