OpenAI Codex Workshop: Hands-On Training for Engineering Teams

OpenAI Codex is not one tool but a family of surfaces: an open-source command-line agent, an IDE extension, and a cloud agent inside ChatGPT that can work on several tasks in parallel and hand back changes for review. A Codex workshop has to teach when to use which surface and how to set a repository up so that all of them produce reviewable, correct work. This guide covers the decisions, the labs and the traps.

Local or cloud: the first thing to teach

Plan availability, limits and some features differ between ChatGPT plans and change over time, so confirm the current details with your account admin and the OpenAI developer documentation before the session.

Lab sequence for a one-day Codex workshop

1. Write AGENTS.md. Codex reads AGENTS.md files for repository instructions. Teams add setup and test commands, code style, directory conventions and rules such as "never edit generated clients." Nested AGENTS.md files can give subdirectories their own guidance.
2. Local bug fix with the CLI. Each pair reproduces a real bug with a failing test, asks Codex to fix it with approvals switched on, and reviews every command it wants to run.
3. Configure the cloud environment. Get dependencies installing and the test suite running in the cloud setup. This is where most teams lose time, so it deserves a dedicated lab.
4. Parallel delegation. Each participant writes two or three ticket-quality task descriptions and assigns them to the cloud agent. While those run, the group discusses review criteria.
5. Review agent pull requests. Pairs review each other's agent-produced changes, request revisions where needed, and decide what merges.
6. Retrospective and AGENTS.md update. Every repeated mistake becomes a line in AGENTS.md.

Writing tasks Codex can complete unattended

Asynchronous agents cannot ask quick follow-up questions the way a colleague at the next desk can. Teach a task template and use it for every lab:

• Outcome: what should be true when the task is done.
• Location: the files, modules or endpoints involved.
• Constraints: what must not change (public APIs, database schema, dependencies).
• Verification: which tests must pass, and which new test should exist.
• Out of scope: tempting adjacent changes the agent should leave alone.

A useful rule of thumb from workshops: if you would need to explain the task in a meeting to a new colleague, it is not ready for a cloud agent.

Five pitfalls that show up in almost every Codex workshop

1. The environment cannot run tests. The agent then "verifies" by reading code, which is much weaker. Fix the setup before delegating anything important.
2. Tasks are too large. A request like "migrate auth to the new provider" returns a sprawling diff. Split by layer or by endpoint.
3. Review fatigue. Parallel agents make it easy to generate more pull requests than anyone can review well. Limit open agent PRs per person.
4. AGENTS.md as a dumping ground. Hundreds of lines of generic advice hide the few rules that matter. Prune regularly.
5. Unclear ownership. Every agent pull request needs a named human owner who is accountable for merging it.

Teams using Codex alongside other agents

Many teams do not standardize on a single vendor. Some developers prefer Claude Code in the terminal, others use Cursor, and the organization may also have GitHub Copilot. The good news is that AGENTS.md is read by a growing number of tools, so one well-maintained instruction file can serve several agents. In mixed teams, the workshop should cover how instruction files, review rules and security policies stay consistent across tools. The comparison of Claude Code vs. OpenAI Codex and the hands-on Codex CLI guide help with that decision.

How to tell the workshop worked

A few weeks after the session, look at the repository rather than at feedback forms. Has AGENTS.md been edited since the workshop, and by more than one person? Are agent-produced pull requests small, linked to a ticket and reviewed by a named owner? Does the cloud environment still run the full test suite, or has it quietly broken after a dependency update? Are developers choosing between local and cloud modes deliberately, or defaulting to one out of habit? Honest answers to these questions tell you whether to schedule a follow-up session, invest in the environment setup, or tighten review rules.

What to prepare before the workshop

• Confirm accounts and plan access for every participant, and connect the code host to the cloud agent if you plan to use it.
• Clarify which repositories may be processed by OpenAI's services under your data policies.
• Pick six to ten real, small tickets with clear acceptance criteria.
• Make sure the project's tests can run from a single command.

neooutput offers live workshops and team trainings that cover Codex next to other coding agents, with labs on participants' own tasks. If you want help deciding which format suits your team, you can book a free intro call.