AI glossary · Advanced · Agents · Safety & alignment
Auto-review
Auto-review is a feature in which, when OpenAI's coding agent Codex tries to do something that goes outside the sandbox, a separate 'reviewer agent' decides whether to approve it instead of a person. It does not expand permissions; it only changes who reviews requests that already require approval.
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Perception & action
- Fields
- Agents, Safety & alignment
- Added
- 2026-10-16
How it differs
A coding agent usually reads and edits files inside a sandbox (an isolated workspace). When it needs to cross the sandbox boundary, such as accessing the network or writing files outside the allowed scope, it asks a person for approval, and the work stops until the person responds. Auto-review sends this approval request to a separate reviewer agent instead of a person. The main agent still works within the same sandbox, the same approval policy and the same network and file restrictions. The only thing that changes is 'who reviews the approval request'.
Human approval vs. auto-review
Human approvalExisting approach
- A person judges each request directly
- Work stops when the person is away
Auto-reviewApprove for me
- The reviewer agent decides and records its reasoning
- If denied, finds a safer path or asks the user
- Permission scope stays the same
공통 Only actions that cross the sandbox boundary are reviewed
How it works
The main agent works in a restricted mode such as read-only or write access to the working folder. When it needs an action that crosses the boundary, it requests approval, and if auto-review is on, the request goes to the reviewer agent. The reviewer agent decides whether the action can be executed and returns its reasoning. If approved, the work continues; if denied, the main agent is instructed to find a clearly safer alternative or stop and ask the user.
How to turn it on
In the ChatGPT desktop app, choose 'Approve for me' from the permissions menu below the input box. In the CLI or local Codex configuration, set approvals_reviewer to auto_review and set the approval policy to one that asks a person, such as on-request. Changing the approval policy alone does not change the reviewer. If the approval policy is never (do not ask), there are no requests to review, so auto-review has no effect. Settings set by an organization administrator take precedence over personal settings.
Things to know
OpenAI's documentation describes Auto-review as 'changing the reviewer, not granting permissions'. It does not add writable folders, open the network, or weaken protected paths. Still, because an AI makes the final decision, the possibility remains that the reviewer agent wrongly approves a risky request. The desktop app shows a warning when you try to give the security model Full Access and recommends 'Approve for me' instead. When working with critical systems, it is safer to also keep human-in-the-loop review, where a person checks directly.
Example
Suppose a developer assigns Codex a bug fix, and during the work the agent requests approval for network access to download an external package. Previously, if the developer was away, the work simply stayed stopped. With Auto-review on, the reviewer agent looks at the request and approves or denies it; if denied, the main agent finds a safer method or asks the developer. This site has covered a report that OpenAI is offering this Codex 'Auto-review' feature free to all logged-in ChatGPT users.