Shipping with Codex

Quick Overview

Codex is evolving into a proactive, powerful AI software engineer teammate, demonstrated by its 10x usage increase from August to October and its integration into internal workflows like code reviews and planning, allowing engineers to ship code faster and with more confidence by handling complex tasks and generating necessary documentation and tests.

Key Points: Codex usage saw a massive 10x shift/increase between August and October due to its integration into daily engineering workflows. Codex is viewed as a collaborative partner, async agent, and proactive teammate, capable of tasks like pairing on code and delegating work without explicit prompting. The Codex Agent architecture consists of a 'Harness' and a 'Model' (like GPT-5 Codex), allowing it to work across various environments (IDE, Terminal, GitHub, Cloud, Mobile). OpenAI engineers are now using Codex to generate test suites (like unit tests) and documentation (like design docs) for complex features, significantly accelerating development. Internal metrics show 92% of OpenAI technical staff use Codex daily, 70% more PRs are submitted per week by Codex users, and 100% of PRs are reviewed by Codex. The development process emphasizes iterative feedback, where Codex helps generate code, tests, and documentation, which are then reviewed, leading to faster, higher-quality shipping.

Context: This presentation, titled "Shipping with Codex," features OpenAI employees Tibo Sottiaux and Aaron Friel discussing the evolution, adoption, and practical application of Codex within OpenAI's own engineering processes. The focus is on how Codex has moved beyond simple code completion to become an essential, proactive AI teammate that accelerates development cycles, particularly in areas like code review and feature planning.

Detailed Analysis

The presentation details the rapid evolution of Codex, positioning it as an AI software engineer teammate rather than just a junior engineer. Tibo Sottiaux highlighted a massive 10x usage increase between August and October, driven by its integration across IDEs, terminals, GitHub, and the cloud. He explained the Codex Agent architecture, which separates the 'Harness' (tools/workflow) from the 'Model' (like GPT-5 Codex). This agent structure allows Codex to perform complex tasks like generating code, test suites, and documentation based on context, as demonstrated by generating a comprehensive design document for a new weather app feature. Aaron Friel followed up by showcasing how Codex is used internally for code reviews, noting that 92% of technical staff use it daily, leading to 70% more PRs submitted weekly and 100% of PRs receiving review suggestions from Codex. He demonstrated using Codex to plan, implement, and ship a feature, including generating tests and documentation, emphasizing that this process is now faster and results in higher-quality code due to the AI's ability to handle complex tasks and provide iterative feedback, effectively acting as a senior engineer.