What is the new Auto-review mode in OpenAI Codex?

Auto-review is a new operational mode for the Codex agentic coding platform designed to enable longer autonomous workflows. It reduces the need for constant manual human approvals by delegating the review process to a specialized secondary agent. This allows the system to maintain progress on complex engineering tasks without frequent interruptions.

How does the OpenAI Codex Auto-review mechanism handle safety?

The system uses a multi-agent architecture where a primary agent performs the coding work while a separate reviewer agent monitors the process. This secondary agent specifically identifies and checks higher-risk steps in context before they are executed. If the step is deemed safe, the automation continues; otherwise, it may pause for human intervention.

What types of engineering tasks is Codex Auto-review designed for?

Auto-review is specifically built to help Codex navigate long-running engineering tasks and complex automations. This includes moving through software builds, executing extensive test suites, and performing other multi-step development workflows. By automating the approval loop, the system can handle these processes more efficiently than traditional modes that require a human to manually approve every individual action.

OpenAI Launches Auto-review to Enable Unattended Codex Agent Workflows

Agentic Coding

GPT

AI Safety

AI Agent

Published Apr 24, 2026

OpenAI launched Auto-review for Codex, a new mode that reduces manual human approvals during autonomous tasks. A separate agent evaluates higher-risk steps in context before execution, allowing the primary agent to continue through tests and builds. This architecture maintains safety while increasing the duration of unattended runs.

This update follows OpenAI's push for secure sandboxes to enable longer autonomous runs. As agents gain computer use capabilities and manage long-term tasks, constant permission requests stall progress. This shift mirrors the per-action safety classification recently introduced by competitors to streamline autonomous workflows.

You can now use Auto-review to keep Codex moving through complex automations, such as repository-wide migrations or multi-step CI/CD debugging, without manual oversight. The feature is designed for long-running tasks where human-in-the-loop requirements previously caused delays. Access is available through the Codex app and CLI.

Frequently asked questions

What is the new Auto-review mode in OpenAI Codex?: Auto-review is a new operational mode for the Codex agentic coding platform designed to enable longer autonomous workflows. It reduces the need for constant manual human approvals by delegating the review process to a specialized secondary agent. This allows the system to maintain progress on complex engineering tasks without frequent interruptions.
How does the OpenAI Codex Auto-review mechanism handle safety?: The system uses a multi-agent architecture where a primary agent performs the coding work while a separate reviewer agent monitors the process. This secondary agent specifically identifies and checks higher-risk steps in context before they are executed. If the step is deemed safe, the automation continues; otherwise, it may pause for human intervention.
What types of engineering tasks is Codex Auto-review designed for?: Auto-review is specifically built to help Codex navigate long-running engineering tasks and complex automations. This includes moving through software builds, executing extensive test suites, and performing other multi-step development workflows. By automating the approval loop, the system can handle these processes more efficiently than traditional modes that require a human to manually approve every individual action.