Codex: the relentless teammate reviewing every OpenAI PR

misceventcodexagentspull-requestssoftware-engineering
tl;dr: I joined OpenAI's Vibe Engineering with Codex event yesterday. Codex as a teammate rewriting a complex project in Rust and reviewing every OpenAI PR really caught my attention. If that sounds interesting, watch the replay and check the event resources I collected!

I really loved yesterday's event, Vibe Engineering with OpenAI's Codex ↗ by the OpenAI team.

Vibe Engineering with OpenAI Codex - live AI automation event featuring Romain Huet, Aaron Friel, Chris Nicholson, and Natalie Cone discussing practical LLM workflows, coding assistants, and real-world AI developer tools

While Codex (OpenAI's coding agent) was running the heavy task of rewriting bazel-diff ↗ in Rust, Romain ↗, Friel ↗ and Chris ↗ from OpenAI shared insightful information about the impact of Codex inside and outside OpenAI.

At the end, they were joined by the enthusiastic Natalie ↗ for a Q&A, with spot on questions like:

  • "Is Codex good for beginners in coding? How does Codex support beginners?"

  • "What would you recommend software engineers do to stay sharp with coding skills when tools like Codex write a lot of code?"

  • etc.

What stands out?

100% of PRs at OpenAI get reviewed by Codex. Period. Simple to picture.

The what does not change: merged code is safe and correct. But the how changes. Codex. Your new relentless teammate.

This relates perfectly to how Romain talks to skeptics about these tools. He avoids abstract arguments and just tells them to try Codex on code reviews ↗: a specific, repetitive task in any software workflow. Easy to understand. Scoped. Then see how Codex can help.

Take a look at things like code reviews. Like, use exactly the tools you have today, but turn on Code Review and see what happens.

—Romain Huet, Head of Developer Experience @ OpenAI

Romain also noticed that curious founders barely have backlogs now. When ideas pop up, they just ask Codex to try them, often in several ways. They keep the best, or discard it all. It's cheap, it's fast. Why keep a backlog?

This ties into Friel's engineer transformation:

For example, with this demo I'm doing today, I will have agents take these tasks on multiple times and I'll compare the results. I will ask them to take different approaches.

—Aaron Friel, Member of Technical Staff @ OpenAI

Besides Codex's insights, I must add that the conversation was so smooth, with Chris's questions shaping a real story:

  • "Are we using Codex to build Codex?"

  • "What is the bottleneck that people can get better at to maximize what these tools can offer?"

  • etc.

Want to watch the replay? 👉 Vibe Engineering with OpenAI's Codex ↗

Thanks Romain, Chris, Friel, and Natalie for the event. I learned a lot.

That's all I have for today! Talk soon 👋

References

Recent posts
featuredHow I crafted TL;DRs with LLMs and modernized my blog (part 1)
See how I crafted story-like tl;drs for my posts with LLMs
prompt
How I use LLMs and Firecrawl to dig into company culture
Learn how I use LLMs to decode company culture from their blogs
code
Did you ask AI?
It's time to build a "just ask AI" habit
Curious about the tools I use?
Eden
My ChatGPT client for the Emacs text editor.
One.el
My minimal static site generator.
Free course
Build a ChatGPT client in Emacs Lisp.