Codex ChatGPT Mobile App Turns Phone Into Approval Layer for Long-Running Agents

OpenAI placed Codex inside the ChatGPT mobile app this week. The move gives users a phone-based control surface for agents that continue running on laptops, Mac minis, or remote servers.

The general availability release adds device pairing, push notifications, goal setting, side chats, file preview, and inline review comments. Users start work on the phone, review results, steer execution, and approve next steps while the actual agent stays on the primary machine.

This setup changes how knowledge workers supervise long tasks. The phone becomes the decision layer instead of another full workspace.

Mobile GA Release Adds Direct Agent Controls

OpenAI announced Codex reached general availability inside the ChatGPT mobile app, as detailed in its official announcement at https://openai.com/blog/chatgpt-mobile-codex-ga/. The update includes one-to-one device pairing that links a phone to a specific computer.

Once paired, the phone receives notifications when an agent finishes a step or needs input. Users can set goals, open side chats for clarification, preview files, and leave inline comments without opening a laptop. For example, a software engineer might tap “approve” on an automated pull request notification during a commute, while a market researcher could open a side chat to redirect an ongoing data-analysis agent between meetings. A financial analyst could preview a quarterly forecast file and add an inline comment requesting variance adjustments, and a corporate lawyer could review draft contract clauses on the phone then approve the next review cycle - both without returning to a desk. Codex itself keeps running on the paired machine. The mobile app only handles supervision tasks. This separation addresses the practical limit that long-running agents cannot stay active on a phone for hours or days.

The change matches how many developers already work. They leave heavy code or research jobs on a desktop while they move between meetings or travel.

Phone Supervision Changes Who Carries the Load

The release puts pressure on teams that still treat agents as single-device tools. Knowledge workers who spend time away from desks now gain a way to stay in the loop without constant laptop access.

Before this update, most agent sessions required the primary device to stay open. Interruptions or travel broke the workflow. The new pairing layer keeps the agent session alive and routes only decisions to the phone.

This shift favors users who already rely on persistent context. Agents that pull from past meetings, documents, and decisions can now receive approvals from any location without losing that grounding.

Device Pairing Creates a Clear Separation of Labor

Codex treats the paired computer as the execution environment and the phone as the review surface. The agent writes code, runs tests, or gathers research on the desktop. The user only approves or redirects at checkpoints.

Device pairing uses secure local tokens rather than cloud relays for the main session. This design reduces data exposure while still allowing quick push notifications.

The pattern mirrors how teams already divide work between primary workstations and mobile oversight. The phone handles short, high-value actions such as approving a pull request or confirming a research direction.

Limits of Mobile Context Create Real Tradeoffs

Mobile supervision still depends on the quality of context passed from the paired machine. If the agent has not captured relevant files or meeting notes, phone approvals risk operating on incomplete information.

As reported by Bloomberg on early enterprise agent pilots, reviewers found complex multi-file refactors - such as interdependent module updates spanning several directories - difficult to judge on a small screen. Preview features help, yet full diff reviews often require the desktop view.

These constraints show that the phone serves best as an approval layer rather than a replacement workspace. Teams that treat it as a complete substitute may encounter missed details or slower iterations.

Persistent Agents Need Grounded Approval Flows

The Codex update highlights a broader requirement for agent workflows: continuous context must travel with the task. Agents that run for hours need access to the same documents, notes, and prior decisions that informed the original goal.

remio addresses part of this need by keeping a five-level memory system that connects meetings, files, and web captures on the user's device. When a Codex agent requests approval, the reviewer can pull relevant history from the same knowledge base instead of reconstructing it from chat logs.

The combination lets knowledge workers approve steps on the phone while the underlying agent still operates against accurate, accumulated context. This approach reduces the repeated explanations that generic agents demand each session.

Next Milestones Will Test Adoption Depth

Teams will watch three signals over the next three months. First, whether OpenAI expands the paired-device list beyond single computers to shared work servers. Second, whether usage analytics show agents completing more steps between mobile approvals. Third, whether competing agent tools add similar phone oversight features.

Each signal will indicate whether the approval-layer model moves from pilot feature to standard workflow pattern.