Codex runs OpenAI workflows through Runme

An OpenAI engineer now runs model evaluations by writing a goal into a Runme notebook and letting Codex plan the work, wait for sign-off, run it and write up what happened. The notebook that comes out keeps the commands, their output, the decisions and the dead ends, and the next run starts from it. The setup is described in a post on the OpenAI developer blog published August 26.
The author describes most of a career spent “turning a crank.” On OpenAI’s cloud infrastructure team, that meant a week per batch of new Kubernetes clusters, fighting private links, quota and Terraform, and then starting on the next batch. On the API team it meant running evaluations against each new model and working through problems with graders, quota, configuration and PyTorch, then doing it all again for the next model. The plan now is to skip building a separate automation for every task. Instead, Runme collects and curates the context around a workflow, keeps review and approval boundaries in place, and feeds what earlier runs learned into later Codex runs.
An evaluation starts as a notebook with a short outline: run the evaluation against the current model, review a previous run to understand the workflow, write a detailed plan in the notebook, wait for approval before beginning, and document the commands, their output and how to interpret the results. Codex then gets a one-line instruction:
Read the goal cell in the Runme notebook open in the browser. Treat it as the goal, write your plan in the notebook, and wait for my approval before starting.
From there Codex reads the notebook and updates it as the work moves. The engineer reviews the plan and revises it where needed, and often the useful input is a choice between alternatives: which evaluation system to use, whether to provision new infrastructure, or whether existing resources will do. During the run the author monitors progress, sometimes from a phone, and nudges Codex when it gets stuck. If a development environment can’t be provisioned because a quota is exhausted, the nudge might be to reuse an existing environment or look for another approved option. Before closing out, the two capture the decisions that would otherwise vanish into the chat: why one option won, which approach is now preferred, and what to do differently next time.
Runme itself is an open-source web application for building notebooks with Codex. Like Jupyter and Colab, it handles Markdown, code cells and HTML, so one document can hold instructions, commands, results, tables and charts. Notebooks save straight to Google Drive, which gives colleagues a familiar place to find them without another document repository. For every notebook Runme also writes a companion Markdown index named *.index.md. Drive can index that file, so an agent looking for examples, operational context or the result of an earlier run can find previous notebooks.
The project’s GitHub repository describes Runme as a tool for making runbooks runnable: it executes shell commands or code inside fenced blocks of Markdown files and, like a terminal session, keeps environment variables between cells. Runbooks stay plain Markdown. The VS Code extension, CLI and Docker builds come from the same project, so the Codex notebook adds an agent to a format already built around step-by-step operations.
Codex talks to Runme through WebMCP. When the app loads, it registers tools in the browser that an agent can call to read instructions for working with Runme and its notebooks, run bounded JavaScript programs that read or update notebook content, and read the app’s documentation. Because Runme is a client-side app served as a static website, a traditional MCP endpoint would have meant adding a server just for that, with extra infrastructure, operational overhead and a change in where notebook data is handled. WebMCP exposes the same capabilities from the page instead, and OpenAI has switched on WebMCP in the ChatGPT desktop app’s built-in browser, where Codex can pick up a site’s tools.
The argument for all this is about context. Much of what would help an agent already exists in day-to-day work but is scattered across terminal history, Slack, runbooks, documents and dashboards. “The hard part isn’t proving that documentation is useful; it’s making the documentation cheap enough to create while the work is happening,” the author writes. Runme puts intent, actions, decisions and outcomes in one artifact. A persistent goal keeps Codex on task, and automatic approval review can clear eligible actions without changing existing permission boundaries. The human’s part shrinks to judging when a plan is ready and making the consequential calls; Codex does the repetitive execution and, in the author’s words, records what happened “in more detail than I would have written myself.” Since the notebook is easy to share, that knowledge doesn’t stay stuck in one person’s chat history.
The post sets this against the tooling that was supposed to make operations easier. Cloud infrastructure and Kubernetes brought a sprawling ecosystem, which the author says the CNCF landscape captures well, and solving one problem often means choosing, learning and running the tools for it. The author describes years spent finding “the magic words that make persnickety machines do what I want,” and hopes Codex takes over the repetitive part while leaving the decisions with people.
“Hopefully, I’ll get some of those heartbeats back—and spend them playing with my dogs.”
Up next
Anthropic signs $45B Nscale compute deal
At Nscale’s planned West Virginia data center, Anthropic is reserving $45 billion in AI cloud power for six years. Nvidia’s Vera Rubin chips will come online late next year to meet its needs.
Related stories
- OpenAI’s lawyers and recruiters now work through Codex
- Leaked code points to a persistent Codex agent
- OpenAI’s Pando cut monorepo push latency nearly 5x
- Codex Desktop hid 323 GB of crash dumps on one Mac
- Left free to choose, OpenAI’s Dots picked Cloudflare 8 of 8
- Codex keeps coding in the cloud after your laptop closes
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

