I was asked to give a talk on how to make the most out of OpenAI's new Codex coding agent, and it turned into a blog post.

Codex works best when you have a really clean code base, with thorough tests, and well documented best practices. Getting your code base into that kind of shape might seem like a lot of work, but Codex is already great at the kind of mundane tasks needed to get your code into a state where Codex can work well.

Codex has also changed how I think about software engineering. With previous AI tools, I used them to help me write the code, but now my job is to mentor a team of AI agents as they write the code.

Even if you don't write code, it's important to understand what it's like to use Codex, because this kind of "manage a team of AI agents to do your job" is likely the future of all human work.