Oracle University announced two courses about OpenAI Codex on October 22, 2026: Codex Foundations and Agentic Engineering with Codex. The first establishes a controlled individual workflow; the second turns those foundations into a team engineering system, from onboarding to a repository through verification and release preparation.
These are not new Oracle certifications, and the announcement does not promise an exam badge or professional title. They are practical learning paths in Oracle MyLearn for developers, technical leads, platform and DevOps engineers, and test automation specialists. Oracle does not publish total duration, price or access period in the announcement, so check those conditions in your account before enrolling.
The two courses at a glance
| Course | Primary audience | Announced modules | Intended outcome |
|---|---|---|---|
| Codex Foundations | Developers and engineers new to Codex | 5 | Use Codex on a project, control context, review and version changes |
| Agentic Engineering with Codex | Experienced users, technical leads and platform teams | 9 | Organize agentic engineering with plans, tools, isolation, tests and guardrails |
The official Oracle University announcement (opens in a new tab) presents Foundations as the conceptual prerequisite for the advanced course. Codex Foundations (opens in a new tab) and Agentic Engineering with Codex (opens in a new tab) are the official destinations, although their content is delivered through MyLearn and may require a session.
Codex Foundations: learn on an existing project
Foundations is not a feature tour. Oracle puts the learner in a scenario: take over a website left by another developer and work through tickets from a product manager. The exercises teach you to explore unfamiliar code, diagnose a bug, add a feature, review the result and prepare a shippable improvement.
The course uses both the Codex CLI and Codex in the ChatGPT desktop app. Learners need a ChatGPT account with Codex access and a computer where they can install software. Oracle notes that the current lab guides use Windows and PowerShell. Teams on Linux or macOS can still use the concepts, but commands and screenshots may not match their environment.
Published objectives include setup, model choice, prompts, context, AGENTS.md, approval policies, sandbox boundaries, code review, Git, branches and reusable skills. The thread connecting them is control of the work rather than automatic acceptance of an output.
What should you be able to do after Foundations?
A learner should be able to explain:
- what information to provide before requesting a change;
- which files and commands the agent may inspect or run;
- when an approval is required;
- how to review a diff before keeping it;
- which tests prove that the expected behavior works;
- how to commit or discard changes cleanly with Git.
To assess yourself before starting, choose a small non-confidential repository and one reproducible defect. Write down current behavior, expected behavior and a neighboring case that must not change. Locate project instructions and tests. This gives you a better starting point than a vague “fix the project” request.
Agentic Engineering with Codex: from tool to team system
The second course covers the full cycle: understand a repository, plan, build, verify and prepare a release. It targets teams that have moved past experimentation and need agent use to be repeatable, observable and compatible with their responsibilities.
Oracle emphasizes durable project context through AGENTS.md, repository maps and context packs. The goal is to stop every session from reconstructing the same rules conversationally. Durable guidance still needs to be concise, versioned and testable. If it becomes a mixture of preferences, secrets, obsolete procedures and product goals, it creates a new source of risk.
The course adds the conversion of business requirements into reviewed plans and acceptance criteria. That transition is decisive. An agent can propose technically correct code that does not solve the real need. An approved plan separates the decision—what should change and how it will be checked—from execution.
Skills, MCP, subagents and worktrees solve different problems
The advanced course names reusable skills, external tools through Model Context Protocol, multiple agents in parallel and Git worktrees. These mechanisms are not interchangeable.
- A skill packages repeatable knowledge or a procedure.
- An MCP tool exposes an external capability or source under defined permissions.
- A subagent separates a task or analysis perspective.
- A worktree isolates the files and Git history of one change.
Choose the problem before combining them. Multiple agents add coordination overhead to a small fix. Connecting an available tool expands access even when the task does not require it. An outdated skill can repeat a weak practice more efficiently.
OnLevelUp's review of AWS Jam workshops for Kiro and AgentCore provides another hands-on view of specifications, agents and observability. It is not a substitute for Oracle's courses: the platform, scenarios and learning outcomes differ.
Verification and runtime evidence
Oracle places tests and runtime evidence at the center of the advanced path. A plausible answer or clean diff does not prove a change works. Teams must define commands, environments and observations that validate the result.
Build a verification record at three levels:
- Static: types, lint, dependencies, secrets, diff and repository rules.
- Behavioral: unit and integration tests, error paths and regression coverage.
- Runtime: real startup, user path, logs, metrics and operational limits.
Add the identity of the person accepting the result and the rollback conditions. This is not an Oracle exam requirement; it translates the announced goal of producing auditable changes before rollout.
Approvals, sandboxing and human control
Both courses cover approval policies and sandbox boundaries. The correct configuration depends on risk. Reading a public repository, changing a local draft, installing a dependency and publishing to production should not receive the same authority.
For every capability, ask four questions: which data can it access, which action can it perform, what is the maximum impact and how is the operation logged? Approval matters only when the reviewer understands the requested action and has enough context to refuse.
Avoid turning “human in the loop” into a decorative button. Place review before the consequential operation, show the exact target and change, and provide evidence. For repeated low-risk actions, automate narrowly and review the policy rather than clicking without reading.
Which course should you choose?
Choose Codex Foundations if you are new to the CLI, do not yet structure context, sometimes accept a diff without tests, or lack a reliable Git method. Its guided scenario can also help a lead evaluate the tool before writing team standards.
Choose Agentic Engineering with Codex if you can already complete an individual task and now face organizational questions: how to version context, isolate concurrent work, connect a tool, assign responsibilities, verify runtime behavior and enforce guardrails.
Take both in sequence when your organization wants a shared vocabulary. Do not launch a multi-agent program before participants can write acceptance criteria, review a diff or distinguish evidence from explanation.
A four-week practice plan
Week 1: context and a small fix
Install the tool in a non-sensitive environment. Map the repository, read its instructions and correct a limited defect. Have another person verify the diff and test.
Week 2: a feature with an approved plan
Convert one requirement into acceptance criteria, risks and a plan. Start implementation only after review. Compare the outcome with the plan, not only with existing tests.
Week 3: one reusable capability and one external tool
Package a single procedure as a skill and connect only one necessary MCP tool. Document permissions, accessible data, expected failures and the disable procedure.
Week 4: isolation, evidence and release review
Run two independent tasks in separate worktrees. Gather the diff, tests, runtime result and rollback plan. In review, explain limits and decisions rather than merely listing changed files.
What to verify before enrollment
Oracle's October 22 announcement does not state the price, total duration, exercise count, subtitle languages, access period or public credential. Do not infer those details from the length of the video tours or from another Oracle University subscription.
Confirm that your ChatGPT account includes Codex access and that your organization's policy allows the labs. Use a training repository or an explicitly approved project. Do not copy secrets, client code or production data into an exercise.
The two courses are useful because they connect daily usage with team discipline. Their value is not a presumed badge. It is the ability to produce changes that are understandable, tested, isolated and reviewable.
