A team at OpenAI built a fully functional software product with zero lines of manually-written code, relying entirely on Codex agents powered by GPT-5. Over five months, a small team guided agents to generate approximately a million lines of code, automate testing and CI/CD, and maintain documentation—all through structured prompts and agent-driven workflows. The human role shifted from coding to designing environments, specifying intent, and building feedback loops.

Key insights:

• Engineers no longer write code—they define constraints, architecture, and feedback systems. Work focuses on scaffolding, abstractions, and enabling agents.
• Agents handle full development cycles, from bug reproduction to fix implementation, testing, and PR creation. Some runs last over six hours, often unattended.
• Repository structure is optimized for agent legibility: AGENTS.md acts as a table of contents pointing to a versioned, in-repo knowledge base (docs/). This ensures agents access up-to-date, contextual information.
• Architecture is strictly enforced via custom linters and structural tests. Domain layers follow strict dependency rules (e.g., Types → Config → Repo → Service → Runtime → UI), and cross-cutting concerns enter through Providers.
• Human review is optional—most PRs are reviewed and approved by other agents. Humans intervene only when judgment is required.
• Automated garbage collection addresses entropy: recurring Codex tasks detect and fix technical debt, enforce 'golden principles', and open refactoring PRs.

The system enables high throughput (3.5+ PRs per engineer per day) and scales efficiently, but raises open questions about long-term architectural coherence and how best to encode human judgment for compounding returns.