Automatic Code Reviews with OpenAI Codex

OpenAI
OpenAINov 4, 2025

Why It Matters

Automatic, high‑precision code reviews eliminate a major human bottleneck, enabling faster, safer releases and freeing engineers to concentrate on higher‑value work.

Summary

Codeex, OpenAI’s newest automatic code‑review agent built on the GPT‑5 Codex model, was unveiled as a plug‑and‑play teammate that integrates directly with developers’ existing tools and workflows. By enabling a simple toggle in the Codex web settings, teams can have every pull request automatically examined by the model, whether the PR is in draft or ready for review, and even receive feedback in the terminal via a new CLI command.

The demonstration highlighted several technical capabilities that set Codeex apart from traditional static analysis. The model scans the entire repository—not just the diff—to resolve dependencies and understand broader context, runs hypothesis‑driven Python tests, and can be steered with custom instructions via the agents.mmd format. Training emphasized high‑precision bug detection, yielding a markedly lower incorrect‑comment rate than prior generations, and the system is designed to be non‑intrusive, surfacing only actionable findings.

Concrete examples showed Codeex catching a subtle React prop removal bug in a VS Code extension and prompting the developer to fix it directly through a conversational loop. Users can augment reviews with bespoke guidelines, request style‑specific responses, or even ask the agent to auto‑fix identified issues. The tool operates both in the cloud for GitHub PRs and locally via the CLI, allowing developers to vet changes before they ever leave their workstation.

If adopted widely, Codeex promises to alleviate the verification bottleneck that plagues fast‑moving AI and software teams, accelerating release cycles while improving code safety. By automating routine yet error‑prone review tasks, engineering resources can focus on higher‑level design work, potentially reshaping how organizations approach continuous integration and quality assurance.

Original Description

Maja Trębacz and Romain Huet show you how to set up Codex to automatically review new pull requests in GitHub and in the Codex CLI, and discuss the research approach we used to train our models to generate quality code reviews.
Timestamps:
00:00 Intro
01:23 Automatic code review in GitHub
02:34 Codex code review vs static analysis
03:31 Training models for high-signal code reviews
04:18 How OpenAI uses Codex code review
05:57 Customizing code review with AGENTS.md
06:58 Code reviews in the Codex CLI
Try it yourself:
Learn more:
Codex: openai.com/codex

Comments

Want to join the conversation?

Loading comments...