Course overview

Codex is OpenAI’s agentic coding tool. It reads files, runs commands, edits code, and reports back with a summary and a diff. It runs in your terminal, inside the ChatGPT desktop app, inside a code editor, in an OpenAI-managed cloud container, and on GitHub pull requests.

This course is an honest introduction for people who do not write code for a living.

Here is the honest part first. Codex is developer-leaning. Most of what it does is aimed at someone with a code repository open. If your business has no repository and no developer, Codex will not do much for you, and this course will tell you that plainly rather than sell you a subscription.

Here is why the course exists anyway. Plenty of Nantucket businesses do have a repository, or a contractor who keeps one, or a website that someone bills them to maintain. If that is you, then understanding Codex changes what you can ask for and what you can check. You can delegate a scoped task to a cloud container and read the diff that comes back. You can have every pull request reviewed before it merges. You can read the safety settings on a tool your contractor is already running against your code and know whether they are set sensibly.

The course also teaches the safety model, and the safety model is the most transferable thing in it. Codex ships with a sandbox, an approval policy, and network access turned off. Those three ideas apply to every agentic tool you will meet, and knowing how one vendor implements them makes you a better judge of the next one.

This is a Delegate course, the top rung of the Nantucket AI trust ladder. Delegate means you hand over a bounded job and check the result. It does not mean you stop looking. Every lesson here treats the review step as the point of the work.

Every product claim in this course was verified against OpenAI’s official documentation on August 5, 2026. Product surfaces, model names, and plan limits change quickly. The last lesson tells you where to look when they do.

Who this is for

• Business owners who pay someone to maintain a website or an application
• Operations managers who approve technical work without writing it
• Anyone who finished Intro to Claude Code or The Four Surfaces of Claude Code and wants to know what the other vendor is doing
• Technically curious professionals evaluating agentic tools before committing budget
• Consultants who need to explain the tradeoff between two coding agents to a client

You do not need to write code. You do need to be comfortable reading a summary of a change and asking whether it is what you wanted.

What you will be able to do

By the end of this course you will be able to:

• Explain what Codex is and name every surface it runs on
• Describe how delegating a task differs from asking a chat question
• Walk through the Codex task loop from goal to reviewed result
• Tell a Codex execution confirmation apart from a business approval
• Explain the three sandbox modes and the approval policies in plain language
• State what Codex can reach on the network by default, and what changes that
• Explain what AGENTS.md is and what belongs in one
• Name the actions that always require a human in this program
• Keep credentials out of an agent workspace and know what to do when one leaks
• Get access to Codex on the plan you already have, or decide not to
• Scope a first task small enough to be safe and useful
• Choose between Codex, a chat window, and Claude Code for a given job
• Recognize when a product claim you learned here has gone out of date

Assessment

Each module ends with a short quiz. The course ends with a final assessment covering all four modules.

Passing grade: 70 percent. Retakes are unlimited. Your score is not published or shared. The quizzes exist so you can check yourself before you point an agent at something that matters.

Modules

1. What Codex Is (3 lessons) – Codex as a mode rather than a standalone app, the surfaces it runs on, and how delegating a goal differs from chatting.
2. The Task Loop and Supervision (4 lessons) – The loop from goal to reviewed diff, the two separate approval systems, sandbox modes and approval policies, and the actions that always need a person.
3. Guardrails (4 lessons) – Sandbox-only habits and responsible use, keeping credentials out, network access off by default, and AGENTS.md as standing instructions.
4. Deciding and Adopting (3 lessons) – Getting access on your current plan, scoping a first safe task, and choosing between Codex, chat, and Claude Code as the products change.

Before this course

Take Intro to Claude Code first if you have not. It teaches the working loop of an agentic coding tool without the vendor-specific detail, and this course assumes you already have that picture. The Four Surfaces of Claude Code is useful context for the configuration lessons in Module 3, though it is not required.

After this course

There is no course after this one yet. The Agent Seat, a capstone on running a supervised agent with its own identity and its own audit trail, is in development.

Frequently asked questions

Do I need to know how to code?

No. You need to be comfortable reading a summary of a change and asking whether it is what you wanted. That review step is the whole point of the course.

Do I have to buy anything to take this course?

No. The course teaches you what Codex is and how to judge it. If you decide to try it, check whether someone in your office already pays for ChatGPT, because Codex is included on every documented ChatGPT plan tier.

Is this useful if my business has no developer?

Partly. The safety ideas transfer to every agentic tool you will meet, and Module 4 will help you decide against adoption with confidence. But if there is no code repository anywhere in your business, Codex itself will not do much for you, and the course says so.

How is this different from the Claude Code courses?

Different vendor, same category of tool. Take Intro to Claude Code first for the general working loop. This course covers OpenAI’s implementation, its specific safety model, and the honest comparison between the two.

Will this course be out of date next year?

Parts of it. Model names, plan prices, usage limits, and install commands change constantly. The sandbox model, the approval model, and the supervision rules change slowly. The final lesson teaches you to tell those apart and points you at the official documentation.

What happens if I fail the quizzes?

Nothing. Passing is 70 percent and retakes are unlimited. Your score is not published or shared with anyone.

Where this course sits

This is a Delegate course on the Nantucket AI trust ladder. Prerequisite: Intro to Claude Code. Related: The Four Surfaces of Claude Code and What Is an AI Agent.