What This Level System Measures
Codex Level Up is a learning structure for expanding the size and risk of work delegated to Codex. It moves from consultation to repository reading, bounded editing, verified implementation, harness design, and parallel operation.
Codex is OpenAI’s coding agent for developer work, available through entry points such as CLI, cloud, and IDE workflows.[1] The Level 0-10 structure in this section is not an official OpenAI ranking. It is this site’s way to explain the learning sequence.
Why Use Levels
Codex can read a repository, propose changes, and support implementation and validation. But if unclear requirements, broad permissions, or vague completion criteria are passed directly to the agent, Codex may fill in decisions that should stay with the human.
The levels separate four concerns.
| Perspective | What the levels clarify |
|---|---|
| Work scope | Whether Codex is consulting, reading, editing, preparing a pull request, or operating a recurring workflow |
| Evidence | Requirements notes, related-file maps, diffs, test results, browser checks, and audit logs |
| Permission | Reading, editing, command execution, external services, and operations that need approval |
| Systemization | Moving repeated behavior into AGENTS.md, Skills, Hooks, Subagents, and validation scripts |
This lets the reader check both “what Codex is allowed to do” and “what the human must still decide.”
The Level 0-10 Map
| Stage | Level | Title | Main goal |
|---|---|---|---|
| Consultation | Level 0 | Chat Advisor | Separate goals, exclusions, and open questions before opening the repository |
| Reading | Level 1 | Repository Reader | Produce a related-file map and change plan before implementation |
| Bounded editing | Level 2-3 | Focused Editor / Verified Implementer | Move from a small diff to verified multi-page implementation |
| Context capture | Level 4-5 | Context Engineer / GitHub Collaborator | Use AGENTS.md, issues, branches, and pull requests to preserve working conditions |
| Harness design | Level 6 | Harness Builder | Turn permissions, prohibitions, validation, and approval rules into operating policy |
| Expanded operation | Level 7-8 | Tool Operator / Parallel Orchestrator | Use browser checks, connectors, and parallel tasks with evidence beyond code |
| Continuous operation | Level 9-10 | Workflow Operator / Agent Platform Architect | Design recurring validation, triage, and an operating model for multiple agents |
Level 6 is the bridge in this section. Once AGENTS.md, approval conditions, and validation commands are explicit, the path connects naturally to Codex Harness, where Rules, Skills, Subagents, and Hooks are designed in more detail. Codex supports custom instructions through AGENTS.md, so repository-level working contracts are a useful design unit.[2]
What You Learn In This Section
This section uses a Next.js personal portfolio site as the running example and expands what Codex is allowed to handle one level at a time.
- Levels 0-1 separate pre-implementation consultation from repository reading.
- Levels 2-3 practice bounded editing and verified implementation.
- Levels 4-5 preserve working conditions through
AGENTS.md, issues, and pull requests. - Level 6 turns permissions, prohibitions, validation, and approval conditions into a harness.
- Levels 7-10 expand into browser checks, parallelization, recurring operation, and multi-agent design.
Each level names not only the deliverable but also the evidence used to judge completion. That matters when reasoning about Codex security and permission boundaries.[3]
Target End State
The final goal is not to let Codex execute everything automatically. The goal is to design objectives, scope, approval conditions, stop conditions, and validation methods so Codex can implement, verify, and report work inside those boundaries.
Near Level 10, the human role shifts from asking Codex for help to designing the operating system in which Codex can work safely. Speed matters, but reviewable diffs, evidence, least privilege, and stop procedures matter just as much.
What To Read Next
After this overview, start with Level 0 Practice. If you want to study operating design first, read Codex Harness after Level 6 to connect the level path with harness architecture.
References
- OpenAI, Codex documentation
- OpenAI, Custom instructions with AGENTS.md
- OpenAI, Codex security
For the latest releases and updates, check the official website and official documentation.