Where the Neophyte sees only the difficulties and the Initiate only the advantages in any great enterprise, the hero sees both.
— Helbrecht, White Dwarf 313 (UK), p. 79
Switching to Codex gave me some of my life back. A permission setting deserves a surprising amount of the credit.
With my earlier Claude setup, I was always doing the yes-no job. Approve. Read the next command. Approve again. I had an agent to do the work, and a new job supervising it.
And until Opus 4.8 came out, I was reading through the diffs too. In both Claude and Codex. The models just weren’t good enough for the harder work I was doing. They got plenty of busywork done. But complex implementations and advanced data structures still needed me to work out the solution and guide the model through it. The supervision went beyond clicking yes.
I had reasons to worry. In my Claude sessions, attempts to fix bugs or build features had involved stashing work, resetting Git state, and throwing away important changes. Those experiences made me reluctant to look away.
--dangerously-skip-permissions did not inspire confidence. The name alone was enough to keep me in my chair.
Me enjoying all the free time my autonomous agent created.
Then I tried Approve for me in Codex.
The useful distinction is between the sandbox and approvals. The sandbox limits where commands can write and what they can access. Approval decides whether a requested exception can proceed. OpenAI documents these as separate controls.
In the usual workspace setup, the choices look like this:
| Mode | What happens |
|---|---|
| Ask for approval | Routine work stays inside the sandbox. Eligible exceptions come to me. |
| Approve for me | The sandbox stays. Eligible exceptions go to an automatic reviewer. |
| Full access | The sandbox restrictions and approval prompts are removed. |
The reviewer gets the proposed action and relevant conversation context. It decides whether to allow it. After a denial, the agent must find a safer approach or ask me. Actions already permitted inside the sandbox don’t need that extra review.
The reviewer is a separate model. The Codex source names codex-auto-review as the default reviewer and gpt-5.6-luna for API-key use. The source leaves the size and price of codex-auto-review unspecified.
In the CLI, the shortcut is:
codex --approve-for-meMy installed CLI describes that as automatic approval review with the workspace-write sandbox. In the app, select Approve for me in the permissions control when available.
My local setup also gives Codex a dedicated temp directory. The configuration is here, including how the named permission profile keeps automatic review enabled.
Claude also documents its own auto mode, with a classifier reviewing actions. In my setup, Codex made this practical enough that I stopped hovering over the permissions dialog.
I still review the finished work, including the tools the agent chose along the way. One habit I noticed was reaching for Python for even tiny tasks, which I explored here.
The constant yes-no duty eased up. I could give the agent a task and actually leave it working. That was the revolutionary bit for me.