MatIntellect

👀 OpenAI rolled out a Codex update that changes the very concept of working with an AI agent. The agent now literally sits down at your Windows computer - sees the screen, moves the mouse, clicks through interfaces and types text - while you control it from your smartphone

How computer use works on Windows:
On May 29 OpenAI added computer use support for Codex on Windows 11 - a first for this platform, on macOS the feature already existed. Codex (built on GPT-5.5) sees the screen through screenshots, understands the state of interfaces and takes actions: clicks buttons, fills in fields, navigates between apps. Tasks you can hand off: testing apps and reproducing bugs through the user's eyes, moving data between programs, automating multi-step workflows without code. Key nuance: on Windows the agent works in the foreground - it takes over the whole desktop. You can't work in parallel - you hand the machine over to the agent. On macOS it runs in the background: things are gentler there

What the agent can do on your computer:
💻 Sees the screen in real time - analyzes screenshots, understands what's open and in what state, builds the task context
💻 Controls the mouse and keyboard - clicks buttons, drags elements, fills in forms, types text
💻 Works across apps - moves data, switches between programs, tests and reproduces bugs in a specific environment

How to control it from your phone:
⚡️ Connect via QR code - Codex Desktop generates a code, you scan it in ChatGPT on iOS or Android - your phone becomes a remote
⚡️ Launch tasks remotely - you type what needs doing from your phone, the agent starts working on the PC, you go about your day
⚡️ Talk to the agent along the way - it asks clarifying questions, sends updates, you approve the next steps - you're not tied to your desk

You can run several parallel threads on one PC and monitor them all from your phone. Security: the agent runs inside Windows Sandbox with limited access to files and the network. What the agent CAN'T do: perform actions that need admin rights, confirm system security prompts, work with terminal apps. At launch it's not available in the EEA, the UK and Switzerland

💼 First launch: Codex Desktop → connection menu → QR code → scan in ChatGPT with the same account → authorization

What I do while the agents work

I already have a practical answer to that question. A social media post used to take me about an hour: find a topic, write it, check the facts, pick an image and rewrite it for each platform. Now the content factory takes it to a finished preview in about 5 minutes, and all I do is pick the topic and press Publish. That has produced 84 posts, none without my approval

I no longer write up client calls by hand. I send the recording to a chat, and a few minutes later I get a transcript, a summary, tasks with deadlines and an updated CRM card. About 40 calls with 15 clients have gone through the system. In the morning the marketer sends a brief on what worked yesterday and what to post today: 80 of them in three months

The freed-up time goes to what an agent won't do for me: talking to clients, making decisions, filming live videos and new projects. The agent takes the routine, and I keep the work that made starting a business worthwhile

The main condition

To lie on the couch with your phone in peace, you need to trust not the model but the rules around it. For me, anything risky, paid or public waits for my yes

🙃 Now the setup literally looks like this: me on the couch with my phone, the agent at the computer doing the work. This used to be a metaphor from articles about "the future of AI". Now it's a QR code, ChatGPT and a Codex window on the desktop. The main question is no longer "can the agent control a computer" - it's "what am I going to do while it works"

Share:

No comments yet

Leave a Comment

Fields marked with an asterisk (*) are required