Agentic Coding Workshops
OpenAI Codex Workshop: Training for CLI, IDE and Cloud Tasks
An OpenAI Codex workshop helps teams use Codex across its surfaces: the CLI in the terminal, the IDE extension, and cloud tasks in ChatGPT that work in parallel and open pull requests. The key skills are writing a good AGENTS.md, scoping tasks so they can run unattended, and reviewing the resulting pull requests rigorously.
What is OpenAI Codex today?
OpenAI Codex is a coding agent available as a command-line tool, an IDE extension and a cloud agent inside ChatGPT that can work on tasks in parallel and propose pull requests.
This multi-surface design shapes how teams use it. Interactive work happens in the CLI or IDE, where an engineer steers the agent step by step. Delegated work happens in the cloud, where several well-scoped tasks can run at the same time in isolated environments and come back as pull requests to review. A workshop should practice both modes and, crucially, the judgment of which tasks belong where.
Because Codex spans several surfaces, teams also need conventions for where work starts and how it is tracked. A common pattern is to launch delegated tasks from tickets, link resulting pull requests back to them and keep interactive CLI sessions for exploration, debugging and work that needs frequent human decisions.
AGENTS.md: the foundation of good Codex results
AGENTS.md is a plain Markdown file in the repository that tells coding agents how to build, test and work within the project, and Codex reads it automatically.
Teams that invest an hour in a good AGENTS.md see noticeably more reliable output. In the workshop we draft it together and keep it focused.
- Exact commands for installing dependencies, building, running tests and linting.
- Project structure and where key logic lives.
- Coding conventions that differ from defaults, such as error-handling patterns or naming rules.
- What the agent must never do, for example editing generated files or migrations without instruction.
- How to verify a change is complete before proposing it.
Scoping tasks for cloud delegation
Cloud tasks work best when each one is small, independent and verifiable by tests.
Good candidates are bug fixes with a reproduction, adding tests to an under-tested module, small refactorings, dependency updates with clear upgrade notes and documentation improvements. Poor candidates are vague features, tasks requiring product decisions, or anything that depends on services the cloud environment cannot reach. In the workshop, participants practice turning backlog tickets into delegable task descriptions with explicit acceptance criteria, then launch several in parallel and compare the results.
Reviewing Codex pull requests
Parallel agents can produce pull requests faster than a team can review them, so review capacity becomes the real bottleneck.
We teach a structured review approach: read the task and the agent's summary, check the tests first, then the diff, and look specifically for unrequested changes, weakened tests, broad exception handling and new dependencies. It helps to limit how many agent tasks run at once per engineer, and to reject and re-scope rather than patch a pull request that went in the wrong direction.
Sandboxing, approvals and security
Codex offers approval and sandbox settings that control what the agent may do locally, and teams should agree on sensible defaults.
In the CLI, engineers can choose how much autonomy the agent has for file edits and command execution. Our recommendation is to start conservatively, then loosen permissions for trusted, well-understood workflows. Keep secrets out of the repository and environment the agent sees, and treat content from issues, web pages or dependencies as potentially untrusted input. These rules belong in the team's written guidelines, not just in individual habits.
Workshop format with neooutput
Our Codex workshops are one or two days, remote or on-site, in English or German, and run on your repository wherever possible.
Participants leave with a committed AGENTS.md, a task-scoping template and a review checklist. Trainers are AI Code Engineers who use multiple coding agents in client delivery, so we can also explain honestly where Codex shines and where another tool might suit a specific workflow better. A short call helps us check your setup and tailor the exercises.
Frequently asked questions
Is AGENTS.md only for Codex?
No. AGENTS.md is an open convention that several coding agents read. Writing it well benefits any agent your team uses, and it doubles as onboarding documentation for humans.
Should we use the CLI or cloud tasks?
Both. Use the CLI or IDE for interactive, exploratory work and cloud tasks for well-scoped, independent tasks that can be verified by tests. The workshop practices choosing between them.
Can Codex work on private repositories?
Yes, when connected through the appropriate account and permissions. Confirm with your security team which accounts and data settings are approved before connecting company repositories.
How many cloud tasks should one engineer run at once?
Only as many as they can review properly on the same day. Running many tasks in parallel is easy; reviewing the resulting pull requests carefully is the real limit, so start small and increase gradually.
