Codex Remote is GA: the agent works, the human approves

On 25 June, OpenAI moved Codex Remote to general availability. The coding agent can now run on a connected Mac or Windows host while you start, continue, and watch it from the ChatGPT mobile app. The notable part is less the new reach than the division of roles it makes visible. The agent works, the human approves. That split is worth a sober look.

What is Codex Remote?

Codex Remote lets you start or continue work on a connected Mac or Windows host, from the ChatGPT app on your phone. You see the progress and approve individual actions directly from an iPhone or Android device. The actual run stays on the host, the control moves into your pocket. With general availability, the changelog no longer notes a plan restriction.

Why the approval is the real point

The appeal is convenience: you kick off a long run, close the laptop, and check on the move whether the agent is on track. But that moves the place where a decision is made. When approval shifts from the desk to the phone, it shifts to a place with a small screen, little time, and half your attention. An approval is not a swipe, it is taking on responsibility. What you approve, you answer for, whether at the monitor or at the bus stop. The convenient path must not become the unchecked one.

Pairing and access

OpenAI has moved Remote Control to an authenticated one-to-one QR pairing between each iOS or Android device and the host. Before connecting, you must update both the ChatGPT app and the Codex app to the latest versions. Connections established since 8 June 2026 stay paired; older, inactive connections require re-pairing. This is the right direction. If you give an agent access to a real machine, you have to control the pairing before you control the work.

The agent stands up its own infrastructure

There is also a new DigitalOcean plugin. With it, Codex can provision a Droplet, configure SSH access, and connect it as a remote workspace. So the agent procures the very machine it computes on. That saves steps and at the same time enlarges the area where something happens unnoticed: a resource that runs, an access that stays open, a bill that grows. More independence means more reasons to set the limits up front, not fewer.

Reliability lives in the architecture

Already with GPT-5.5 the point held: more autonomy only moves the place where a mistake shows up. Codex Remote does not change what the agent can do, but where you approve it. Reliability does not come from the agent, but from what you build around it. Define where a human signs off, and hold those points even when approval is only a thumb-press away: before the merge, before the deploy, before the migration.

Where Codex Remote earns its place

The progress is real and practical. Let the agent run the long, tedious stretches on the host, and use the app to see progress, not to nod through responsibility on a small screen. Keep your hand on the points where an unnoticed step costs money, code, or trust. The agent works, the human approves. That order is the gain, not the convenience of skipping it.

Sources

• OpenAI: Codex Changelog