Your AI Agent Is Becoming a Pocket Coworker — Here’s How to Manage It
Your AI Agent Is Becoming a Pocket Coworker — Here’s How to Manage It
OpenAI putting Codex into the ChatGPT mobile app sounds, at first, like a developer feature.
It is not.
Yes, Codex is still aimed at coding work. But the behavior shift matters far beyond developers: an AI agent is no longer something you only babysit while sitting at your computer. It can keep working while you step away, ask you questions from your phone, show progress, wait for approval, and let you redirect the work before it goes too far off the rails.
That is the important part.
The future is not “do your whole job from a phone.” Please do not try to run your business from a six-inch glass rectangle while pretending this is enlightenment. The useful shift is simpler: your agent can now behave more like a pocket coworker.
Not magic. Not autonomous genius. A coworker.
One that needs context, boundaries, checkpoints, and the occasional “no, stop doing that.”
What Codex mobile actually signals
OpenAI’s announcement says Codex is rolling out inside the ChatGPT mobile app on iOS and Android. From mobile, users can review outputs, approve commands, change direction, follow screenshots, inspect test results, look at diffs, and move across active work threads. OpenAI also says files and credentials stay on the trusted machine rather than being copied onto the phone. Source: https://openai.com/index/work-with-codex-from-anywhere/
For everyone else, it shows the shape of the coming AI agent workflow: start a task, walk away, answer questions, approve risky steps, redirect when needed, and review the finished work later. Normal apps wait for you. Agents work between your check-ins.
That is why the Hacker News discussion around “Codex is now in the ChatGPT mobile app” drew so much attention. The work thread follows you.
The right mental model: pocket coworker
A lot of AI agent confusion comes from using the wrong metaphor.
If you treat an agent like a magic app, you will either overtrust it or get frustrated when it asks basic questions. If you treat it like a junior coworker with infinite patience, things get easier.
A junior coworker can research, draft, compare, organize, summarize, and prepare options. But they need to know what “good” looks like, what they are allowed to touch, and when to ask first. Your mobile agent is the same.
That pause is not annoying. That pause is the product.
Good mobile-agent tasks vs. bad mobile-agent tasks
Good mobile-agent tasks have safe checkpoints:
Good tasks:
- “Research three scheduling tools and summarize the tradeoffs.”
- “Draft a customer follow-up email based on these notes.”
- “Review this landing page and suggest improvements.”
- “Prepare a Monday priority list from my open tasks.”
Bad tasks can quietly create a mess:
Bad tasks:
- “Send refunds to everyone who complained.”
- “Change my pricing page without asking me.”
- “Reply to angry customers in my voice.”
- “Make account or permission changes without checking first.”
The difference is whether the task has safe stopping points. If it can quietly break trust, money, access, or customer relationships, it needs tighter supervision.
The three questions to ask before delegating longer work
Before you hand an agent a task and walk away, answer three questions.
First: what can it see?
Second: what can it change?
Third: when should it ask first?
These are boring questions. Boring is good. Boring keeps your agent from becoming a very fast intern with scissors.
This is where beginners get stuck. The answer is not “trust the agent more.” The answer is to define the working agreement better.
Try these reverse prompts
Use these with your own agent before delegating longer work from mobile.
Before you start, tell me what you need access to, what you plan to change, and where you will stop to ask for approval.
Work in checkpoints. After each major step, summarize what changed, what you learned, and what decision you need from me next.
If you are uncertain, do not guess silently. Give me two or three options with your recommendation and the tradeoff for each.
Do not make irreversible changes without asking first. Draft, prepare, or preview the change, then wait for my approval.
If I come back later, give me a plain-English status update: what is done, what is blocked, what you need from me, and what you recommend next.
These prompts are not fancy. That is the point. Good agent management usually sounds like clear coworker management.
The operator skill is steering
Mobile agents will tempt people into two bad habits.
One group will micromanage every step and complain that the agent is slow. The other will give vague instructions, disappear, and act shocked when the agent optimizes for the wrong thing.
The useful middle is steering.
Give the agent enough context to move. Set boundaries around risky actions. Let it work between checkpoints. Review summaries. Redirect drift.
That is the new workflow ChatGPT Codex mobile points toward. It is not about coding on your phone. It is about long-running AI work becoming easier to supervise from wherever you are.
For non-technical operators and solopreneurs, this is the skill to build now: learn how to manage AI agents before you need an agent to manage something important.
Because the agent in your pocket is not replacing your judgment.
It is giving your judgment a new place to sit.