OpenAI Built a Coding Agent. Then Its Own Engineers Started Depending on It.

When OpenAI published an internal guide to how its engineers use Codex, it wasn't a product announcement. It was closer to a field report — observations from teams across the company that had quietly built the coding agent into their daily routines.

The document covers Security, Product Engineering, Frontend, API, Infrastructure, and Performance Engineering. What stands out isn't the breadth but the specificity of how people are actually using it.

A site reliability engineer on API Platform pastes stack traces into Codex during on-call incidents to locate authentication flows without needing to know the codebase cold. A backend engineer on ChatGPT Web described letting Codex hunt down every instance of a legacy getUserById() call, replace it with a new service pattern, and open the pull request — work that would have consumed most of a day. On Model Serving, a platform engineer saved 30 minutes of performance tuning with a five-minute prompt. An infrastructure engineer on API Reliability uses it to scan for repeated expensive database calls and sketch out batched alternatives.

The best practices that have emerged are practical rather than polished. Engineers recommend writing prompts like GitHub issues — specific, scoped, with enough context to be actionable. There's an AGENTS.md convention taking hold: a file in each repository that gives Codex standing context about the codebase, so engineers aren't re-explaining the same architecture every session. The task queue doubles as an informal backlog. For gnarly problems, the Best-of-N feature — which generates several candidate solutions simultaneously — has become a go-to.

One product engineer on ChatGPT Enterprise merged four pull requests on a meeting-heavy day. Codex had been working in parallel the whole time.

The document doesn't read like a case study written to sell something. It reads like notes from people who found a tool useful and started pushing it further. That's a different kind of credibility than a launch blog post — and a more honest signal for anyone watching how AI agents are actually landing in professional software development.