OpenAI codes 1 million lines for an internal product

This is how OpenAI shipped a product with over 1 million lines of code, where every line of code has been written by AI. The problem is that AI agents are unreliable by default. They hallucinate file paths, violate module boundaries, and accumulate dead code over time. Raw model output degrades as the code base grows. The solution is harness engineering. A horse is powerful but directionless. The harness channels that power into useful work. We use the same idea here, by structuring the AI’s output The first technique is context engineering. The agent’s entire knowledge lives in the repository as versioned markdown. Chat history is compacted, preferences are stored as config files, and every design decision is an indexed document the agent can retrieve. 2. Architectural Constraints Code follows strict dependency layers. Types -> Config -> Service - > UI Linters enforce this mechanically, so the agent cannot violate module boundaries. 3. Entropy management AI-generated code accumulates bloat. Stale docs, dead functions, and naming drift. Background agents run continuously, scanning and cleaning the codebase. They estimate this was built in 1/10th the time manual coding would have taken. Three engineers built it in five months using Codex agents. They averaged 3. 5 pull requests per engineer per day, across 1,500 merged PRs. Big claims from OpenAI... but without independent verification This reads more like a marketing pitch for Codex than an engineering case study. Thanks for watching. Cheers!