Put it to work.
You've got the picture now: what Codex is, how to set it loose safely, the jobs it does well, and how to keep people in the loop. The last step is the one that actually changes anything, because reading about an agent teaches you nothing until you've watched one work on your own code. This week, you trial it. Here's the calm way to start.
Pick a few well-scoped jobs to trial
Don't reach for a sprawling rewrite to prove the point. Pick a handful of small, bounded jobs with clear done criteria, the kind from lesson three: a checkable result, low blast radius, and no production secrets in sight. Good first candidates for almost any codebase:
- Add test coverage to a module that's light on it.
- A tidy refactor with the suite as your safety net: rename, extract, split a fat file.
- A small bug with clear reproduction steps, so it can write the failing test and fix it.
- Glue code between two services, or a bit of boilerplate that follows an existing pattern.
- A small internal tool or one-off script that saves the team a recurring chore.
Write your shortlist down before you start, and make sure your repo is ready: a project instructions file with how to build, test and lint, and branch protection with a green CI as the backstop. A few real tasks beats a vague intention to "try the agent" every time.
Run the loop: scope, run, review, merge
For each job, the rhythm is the same, and it's quick:
- Scope. Write a tight brief: one clear goal, a definition of done, and the boundaries. If you can't scope it cleanly, break it down first or keep it for yourself.
- Run. Hand it to the agent and let it work in the sandbox or your checkout. Run a couple in parallel if they're independent.
- Review. Read the diff like any contributor's. Is the approach sound? Anything off? Green tests are a start, not a sign-off. This is where your judgement goes in.
- Merge. If it's good, merge it through your normal protected flow. If it came back off, tighten the brief or improve the context and run it again. That feedback is how both of you get better.
That last habit is the compounding bit. Every time you sharpen the brief or add a missing note to your instructions file, the next task lands closer to right. Over a few weeks the agent quietly clears a real slice of the well-shaped work, and your people spend more of their hours on the parts that need them.
The mindset to keep
Codex is leverage on the well-defined work. It is not a replacement for your engineers and it is not a reason to skip review, and the human check is still the thing that keeps what you ship good and safe. Let it carry the bounded, checkable jobs. Keep your hands on design, on the risky paths, and on the decision about what's worth building and what's worth shipping. That balance is where the speed is real and the risk stays low.
Keep the summary handy
To make this stick, grab the free one-page summary: what Codex is, the scoping rules, the sandbox and permissions, the secrets line, the review and CI guardrails, and a space for your first tasks. Print it, stick it by the desk, and you've got the whole course in arm's reach.
A few quick questions to lock it in. No marks recorded, just for you.
Answer all the questions to continue.
That's the course done. Nice work.
You've got the full picture now: what Codex is, how to set it loose safely, the jobs it does well, and how to keep people in the loop. Here's what to do next.
Answer the quick check above to unlock this.
Now put it to work
Turn all this into action with the fill-in workbook: spot where a coding agent helps, scope your first tasks, set the sandbox and permissions, and write your review and secrets policy. Pop your email in and work through it with your team.
You're in.
Your fill-in workbook is ready. Open it below, then type straight into it or print it to work through with your team.
Open your workbookSave your progress
Pop your email in and we'll send you a link to pick up where you left off, on any device. No account needed.
Saved.
Check your inbox for a link to continue on any device.
Keep learning at your own pace.
The other free JDCS courses go broader on AI and automation, and the guides dig into the parts that matter most to how you build.
Want a hand putting an agent to work in your team?
Tell me a little about how your team ships and I'll come back with a plain-English read on where a coding agent would help most, and the guardrails to put around it first. The first conversation is free, and you keep the plan either way.
Got it, thanks.
That's landed with me. I'll read it properly and come back with a plain-English plan, usually within a few hours and always within one business day.