What a Codex course should teach your engineers

The test for any Codex course

Ask one question before you book: does the session run in our repository? A demo project can teach mechanics, while a team repository tests transfer to your own build, conventions and review process. Ask how the course connects the two.

Tool mechanics are only one possible source of friction. It is a seven year old build script, an untested module everyone is scared of, and a review culture that was designed around humans writing two hundred lines a day.

What belongs in the syllabus

• The loop. What the agent can see, when it decides, why it gets stuck. Without this mental model people either over-trust or refuse to use the tool.
• Repository preparation. Instruction files, exact test commands, exclusions, and the CI checks that make agent output verifiable.
• Task shaping. Turning a ticket into a prompt with a finish line. This is where the largest quality difference sits, and it is teachable in an hour.
• Reviewing agent diffs. Different failure modes than human code: plausible but unnecessary changes, silent scope creep, tests written to pass rather than to catch.
• Team rules. Approval modes, which repositories are in scope, what never goes near an agent.

A concrete exercise to request: take a real bug from the team's tracker, have everyone write a prompt for it independently, run all of them, and read the diffs side by side. People argue. That argument is the training.

How long it should take

Duration should follow the learning goals, starting experience and amount of supervised practice. A short session can introduce the tool; consistent independent use requires evidence from subsequent tasks. Compare proposed exercises and follow-up support instead of treating a number of days as a guarantee.

Recorded video courses have a specific weakness here. They cannot show you what happens when the agent hits your build. That moment is where people learn.

When to skip the course

If your test suite does not run reliably on a developer machine, fix that first. Agents work by iterating against a check, so a team with no working check gets fast generation of code nobody can validate. Use a controlled exercise while repairing that verification gap, and do not claim the team is ready for autonomous production changes.

Next step

Before you buy anything, run an internal dry run. Pick three engineers, give them a week with Codex on real tickets, and ask them to write down every point where they got confused or gave up. That list is your syllabus. Bring it to whoever you hire, including us, and refuse to be taught anything that is not on it.

If you want help putting this into practice, talk to us.