OpenAI's team built an internal product over five months with zero manually-written code, using Codex agents to generate approximately one million lines across 1,500 pull requests. The experiment revealed that engineering work shifts from writing code to designing environments, creating feedback loops, and building scaffolding that enables agents to work autonomously. Key learnings include treating documentation as a navigable map rather than a manual, enforcing architectural invariants mechanically, making all context repository-local for agent legibility, and implementing continuous automated refactoring to prevent technical debt accumulation. The system now supports end-to-end feature development where agents can reproduce bugs, implement fixes, validate changes, and merge pull requests with minimal human intervention.
Table of contents
We started with an empty git repositoryRedefining the role of the engineerIncreasing application legibilityWe made repository knowledge the system of recordAgent legibility is the goalEnforcing architecture and tasteThroughput changes the merge philosophyWhat “agent-generated” actually meansIncreasing levels of autonomyEntropy and garbage collectionWhat we’re still learningSort: