JA EN
shiichan

Goodbye, Repetitive Work! How OpenAI Uses Codex Internally

OpenAI Developers ★★★ View Markdown
#openai#ai#devops

Hi, it's me, Shiichan! Today I want to share a clever workflow trick I found on the OpenAI Developers blog.

OpenAI Developers developers.openai.com

What was announced?

OpenAI Developers, OpenAI's blog for developers, published a post about how OpenAI's own engineers cut down on repetitive work. It covers tricks for handing tasks like spinning up Kubernetes clusters on the cloud infrastructure team or running model evals on the API team over to Codex.

The story so far

Until now, these kinds of tasks had to be done by hand, one step at a time, with engineers recalling the exact procedure each time. Repeating the same cluster setup or eval run over and over led to inconsistent steps, small mistakes, and a lot of lost time.

What changes

The post describes a workflow built around Runme, an open-source notebook app. Engineers write a "goal cell" describing what they want to accomplish, then tell Codex to read the notebook, write a plan, and wait for approval. Codex drafts the plan and updates the notebook, a human reviews and adjusts it, and only then is it approved to run. Everything that happens gets recorded right in the notebook, so the next time the same task comes up, there's already a record to build on.

Dive Deep

Runme is an open-source app that supports Markdown, code cells, and HTML. Notebooks can be saved straight to Google Drive, and each notebook comes with a Markdown index file named *.index.md.

Codex can act directly inside a notebook thanks to WebMCP, which exposes browser-side tools. Through it, Codex can read Runme's own usage docs, read and write notebook content, and run a restricted JavaScript program.

The cycle looks like this: run an eval, record the workflow in the notebook, use that record to improve Codex's workflow, and make the next eval run faster.

Wrap-up

  • The OpenAI Developers blog shared how OpenAI's engineers use Codex to automate repetitive internal work.
  • Engineers write a goal in an open-source notebook app called Runme and let Codex plan and execute the task.
  • WebMCP lets Codex read and write notebook content and even run JavaScript.
  • Every run gets recorded, so the automation keeps improving over time.
  • Worth a look for engineers who want to offload routine work — like spinning up clusters or running evals — to Codex or a similar AI agent.

Related posts