A complete assignment reduces review friction
Provide the observed problem, expected behaviour and relevant constraints. A useful assignment also states what counts as evidence: a reproducible test, correctly rendered state or measurable behavioural change. For an initial pilot, choose a bounded task with a known target. Review changes unrelated to the assignment separately.
Sandboxing and approval operate at different levels
The sandbox technically constrains what executed commands can access; the approval policy governs required confirmation. Documented defaults restrict write permissions and network access, but configuration can vary by environment. This does not mean the whole product works offline. Before the pilot, review the effective file, network and tool permissions.
Distinguish cloud setup from agent execution
Codex Cloud setup can use network access to install dependencies. The agent phase has its own network rules. Assess these phases separately for your workflow. Keep required package sources, test data and interfaces explicitly bounded. A test that works only with uncontrolled production access is not yet a suitable implementation path.
Example: an internal form with clear acceptance
A team wants to improve required fields and error messages in an internal form. Codex works on an isolated change, runs the intended tests and provides a reviewable diff. The reviewer checks keyboard use, invalid inputs and existing form paths. The change is adopted only after acceptance. The pilot also measures manual correction and review time.
Assess implementation against the repository
The AI System Check identifies stalled engineering work and existing prerequisites. Repository rules, test commands, permissions and review routes can then be configured concretely. If the build or documentation first needs repair, that effort is made visible. Successful implementation should reduce the team’s workload while preserving maintainability.
Decision matrix
| Check | Before the pilot | Evidence |
|---|---|---|
| Starting state | Branch or worktree with a known starting commit | Diff contains the intended change |
| Permissions | Bound files, commands, network and tools | The permitted workflow works within the boundary |
| Quality | Define acceptance and regression tests | Tests plus professional review |
| Value | Record baseline effort | Compare total time including rework |
Keep it verifiable
Primary sources
- OpenAI: SandboxingSource checked:
- OpenAI: Agent approvals and securitySource checked:
- OpenAI: Agent internet accessSource checked:
- WorktreesSource checked:
- Code reviewSource checked:














