OpenAI Agents API adds automatic approval via auto_review: A review model approves or rejects high-risk operations on behalf of the agent

The OpenAI developer account reposted the Agents API team's weekly update on 10 30, noting that Auto Review brings Codex’s "approve for me" feature to the Agents API. According to official documentation, when a session's review field is set to "auto_review", an OpenAI-hosted review model evaluates operations requiring approval (such as shell commands needing elevated privileges or posing risks, file modifications, expansions of filesystem or network permissions, and MCP tool calls that rules require confirmation for). The model uses conversation context to determine what the operation does, its risks, and whether it aligns with user intent, then automatically approves or rejects it. If not set, the default is "none".

The OpenAI developer account reposted the Agents API team's weekly update on 10 30, noting that Auto Review brings Codex’s "approve for me" feature to the Agents API. According to official documentation, when a session's review field is set to "auto_review", an OpenAI-hosted review model evaluates operations requiring approval (such as shell commands needing elevated privileges or posing risks, file modifications, expansions of filesystem or network permissions, and MCP tool calls that rules require confirmation for). The model uses conversation context to determine what the operation does, its risks, and whether it aligns with user intent, then automatically approves or rejects it. If not set, the default is "none".

Confidence: High (80) · Event time 2026-10-30

Rationale

  • When building agents for long-running autonomous operation, developers typically face a choice between waiting for manual approval at every step or granting broad permissions. Automatic approval offers a middle ground, eliminating the need for developers to write their own approval event handling logic.
  • Setting the value to "none" does not route operations to the application for manual review; instead, most operations execute directly (subject to environment permission constraints). Understanding this distinction is critical for security configuration.

Usage

  • Add review: "auto_review" when creating a session (at the same level as agent or agent_id, not within the agent configuration). You can revert it to "none" using the update session endpoint when tasks are idle. See the "Approvals" page for the Agents API on developers.openai.com.

Risks

  • Tokens used for review incur additional charges; automatic approval is not supported when reusing existing hosted environments (environment_id), and self-hosted environments must first configure sandboxes according to documentation.
  • Other updates mentioned in the same tweet, such as Async Compaction, were not found in the documentation or changelog by the editor and are therefore excluded. The specific launch date for auto_review is not stated in the documentation.

Open · OpenAI Documentation: Agents API Approvals

X @OpenAIDevs

Explore the latest AI Radar updates