An experiment where a team built and shipped a complex software product entirely through AI agents, with zero lines of manually-written code.
Over five months, three (later seven) engineers used OpenAI’s Codex to generate a million-line codebase—including application logic, tests, CI configurations, and documentation. They estimate this was achieved in 1/10th the time of traditional manual coding.
The core philosophy was: “Humans steer, agents execute.”
Key learnings include:
- The Engineer’s Role Changed: Engineers stopped writing code and instead focused on designing environments, specifying intent, and building the tools and scaffolding required for Codex to do reliable work.
- Agent Legibility is Crucial: For the AI to function, the entire context of the project had to live in the repository. Decisions made in Slack or Google Docs were invisible to the agent. The team built custom tools giving Codex access to Chrome DevTools and observability stacks so it could reproduce bugs and validate fixes autonomously.
- Maps over Encyclopedias: Providing the AI with one massive instruction file failed. Instead, they used a short “table of contents” file that pointed the agent to deeper, structured, and mechanically verified documentation.
- Strict Architecture: To prevent the fast-moving AI from creating a messy codebase, the team enforced rigid architectural boundaries using custom linters and tests. The AI had freedom within these boundaries, but could not violate them.
- Automated Garbage Collection: AI naturally replicated existing code patterns, including suboptimal ones (“AI slop”). To combat this, the team deployed background Codex agents to constantly scan for technical debt, enforce “golden principles,” and automatically open refactoring pull requests to keep the code clean.
Ultimately, building software with agents requires immense engineering discipline, but that discipline is applied to the system’s guardrails rather than the code itself.