Codex auto‑review video surfaces
- A YouTube explainer posted April 25 says OpenAI’s Codex now has an auto-review mode that lets the coding agent keep working longer while machine checks reduce some human approval prompts. - OpenAI’s docs say Codex can auto-review pull requests in GitHub, follow repository-specific AGENTS.md review rules, and in Auto mode still stops for network access or edits outside the workspace. - The shift is from one-off code generation to controlled software execution inside sandboxes, approval policies, and audit trails. (developers.openai.com)
Coding agents are moving from writing snippets to running longer jobs, and OpenAI’s Codex is adding more machine review before a human has to click approve. (youtube.com) (developers.openai.com) The basic problem is simple: an agent that can edit files and run commands is useful, but every approval prompt slows it down. OpenAI’s Codex docs describe two controls for that tradeoff — a sandbox that limits what the agent can touch, and an approval policy that decides when it must stop and ask. (developers.openai.com) In Codex’s default local setup, network access is off and write access is typically limited to the active workspace. In the Auto preset, Codex can read files, make edits, and run commands in that working directory automatically. (developers.openai.com) Codex still asks for approval when it needs to edit files outside the workspace or run commands that require network access. OpenAI also says destructive tool calls require approval when the tool advertises that risk. (developers.openai.com) That means “fewer approvals” does not mean “no guardrails.” It means pushing routine work inside a fenced area, then escalating only when the agent tries to cross a boundary such as internet access, broader filesystem access, or destructive actions. (developers.openai.com) The review layer is already visible in GitHub. OpenAI says teams can ask for `@codex review` on a pull request, or turn on automatic reviews so Codex posts review comments whenever a new pull request is opened for review. (developers.openai.com) Those reviews can be tuned with repository instructions. Codex looks for AGENTS.md files and follows “Review guidelines” set by the team, including checks like “Don’t log PII” or “Verify that authentication middleware wraps every route.” (developers.openai.com) OpenAI’s GitHub docs say Codex flags only P0 and P1 issues by default, which narrows the system toward high-severity findings instead of commenting on every small style choice. A one-off prompt can also focus the review on a category like security regressions. (developers.openai.com) The broader change is that code review is becoming part of agent execution, not just a check after the fact. OpenAI’s Codex materials now pitch the product as a tool that can review code, debug failures, automate development tasks, and operate in cloud or local sandboxes. (developers.openai.com) (openai.com) The result is a narrower question than “can the model code.” The live product question is which jobs can run unattended inside a sandbox, and which still need a person at the boundary. (developers.openai.com)
