Overview
-
The Challenge: After Sora's iOS launch exploded in popularity, OpenAI needed to ship Android quickly with only a small prototype and mounting pre-registered users.
-
The Team Approach: Instead of adding engineers and process, a lean team of four used Codex as force multipliers, treating it like a senior engineer needing onboarding and direction.
-
Where Codex Needs Guidance: Cannot infer architecture preferences, product strategy, or internal norms. Struggles with deep architectural judgment and can't observe the app running on actual devices.
-
Where Codex Excels: Reading large codebases, writing test coverage, applying feedback from CI failures, running parallel disposable sessions, and enabling engineers to focus on review rather than writing code.
-
Foundation First: The team built core architecture, authentication, and networking themselves, then wrote representative features as examples for Codex to follow.
-
Planning Before Coding: For non-trivial changes, they first asked Codex to summarize how systems work, refined its understanding, then created implementation plans before generating code.
-
Cross-Platform Translation: The iOS codebase served as living documentation, with Codex translating Swift implementations to Kotlin while preserving semantics.
Takeaways
OpenAI's engineering team built this with Codex, the same tool available to any developer. AI-assisted development increases rather than reduces the need for engineering rigor and systems thinking.
The super powers of tomorrow's software engineer will be deep systems understanding and the ability to work collaboratively with AI over long time horizons.