OpenAI Codex for Research Engineers: First Task to Reviewed Change
On a research team, much of the code an agent touches computes numbers people act on: an eval script, a metric, the function that fills a results table. A change that moves a headline number and still passes its tests is easy to miss. This course trains the check that catches it: record what the code produced before, ask Codex for one bounded change, read the diff, and rerun the baseline. Then you write AGENTS.md, fix the working directory for scripted runs, and check what instructions a run received. You practice on a stand-in, a small Python bill splitter whose outputs you can predict by hand; each lesson then writes the same steps out for an eval repository. The practice tasks were run on Codex CLI 0.155.1 on September 21, 2026. The research-code examples have not been run yet, and each says so. Product claims quote OpenAI's documentation. You need a terminal, Git, Python 3.9 or newer, and Codex through ChatGPT or an OpenAI API key.
ToolChecked against OpenAI Codex CLI 0.155.1 on 2026-09-21.
Rod Rivera
Professor
Desk
Every lesson is free. Lessons marked open need no account; the rest open once you register.
What You'll Learn
- How to set a baseline before an agent changes code that produces numbers, so you can compare afterward
- How to ask Codex for one bounded change and read what it did
- Why your own rerun and a diff you have read are the evidence, and the agent's summary is not
- How to write the same task for an eval script, so a change cannot move a headline number unnoticed
- Which rules an eval repository should write down for an agent, and why each one protects a results table
- How Codex finds AGENTS.md files, and which one wins when two disagree
- How to fix the working directory so a scripted or scheduled run behaves the same from anywhere
- What a model's account of its own instructions shows, and which check settles where they came from
Prerequisites
- Basic Python, enough to read a short function and a unit test
- Helpful, not required: some experience running evaluations or experiments in Python
- A terminal, Git, and Python 3.9 or newer
- A ChatGPT plan that includes Codex, or an OpenAI API key billed at OpenAI's standard API rates
Read the rest of OpenAI Codex for Research Engineers: First Task to Reviewed Change
The first lessons need no account. Sign in free to read the rest and keep your place on any device.
