Give it a full desktop in a VM if you want to, just not direct access to your system.
Also in the news: https://www.whitehouse.gov/presidential-actions/2026/08/expa...
With the recent security related events we already see that it seems to be wide spread practice that many operators avoid the hassle of isolating their agents.
It is just an incredible cultural experience to see how quick stupidity gets accepted as a behavioural default.
A VM would be even better, but people could presumably run this in one if we manage to convince them of higher sandboxing in general ?
Tabs for long-running often-compacted sessions is a difference I've noticed in emergent UX.
A senior engineer in an Open Source project told me "I haven't looked at code in months"
In which case what TF is it you even do ?
1. Much better multi-agent orchestration. The app supervises many agents, with different project views and built-in git worktrees, so agents can work on isolated copies of the same repo. It's easy to jump between tasks and see changes. I recall OpenAI wrote that CLI was not designed around this style of work.
2. The app combines multiple terminals, file previews, PR-review workflows, SSH connections, an integrated browser, and page annotations. For front-end stuff, Codex can run the app in Codex for mockups/prototypes, then iterate inside the app. IMHO I think this is a much faster, more intuitive flow.
3. For long-running and recurring agent workflows, I think they're much easier to manage in the app. You get automations, review queues, persistent conversation context, future schedules, and auto resume for longer running tasks. E.g. CI failure summaries, issue triage, and release briefs.
You can do some of the above, to some degree, with varying levels of success in the CLI, but I think it's miles ahead in the app.
However, from the screenshots and description, I expect that you would have these advantages:
- History. Currently, if you type into the CLI (TUI) your intention and requests are captured in some transcripts, but not in conversational turns. If you want the history of what you typed into the cli/tui you would have to datamine it from transcripts. By contrast, on the screenshots you can see it is a simple conversation that's saved and easy to return to.
- Multiplatform. As per the announcement "Use Codex across ChatGPT, your editor, and the terminal", meaning that you could open the ChatGPT app on your phone and continue to interact with this. While this might not matter to you if you exclusively code while sitting in front of your computer, still, it helps you be able to stay on top of things if you are on mobile.
- Background tasks and autonomous work. It says you can schedule it to do things in the background (issue triage, alert monitoring). At the moment, it would only do this if you've asked either codex or another agent to build a harness for it that does so. (i.e. you could ask Claude to build a harness that calls Codex on some schedule or interval.)
These all seem pretty good and I look forward to starting to use it once it's been more battle-tested.
[1] notably, on the Anthropic side the Opus 5 regressions - I had to downgrade to Opus 4.8 as I found Opus 5 was unreliable. I currently use Opus 4.8 and Codex Sol 5.6.
You can give it any task on your computer and it can automate it. Such as creating a photo on Photoshop, creating 3D models, editing videos, or browsing any website without scripting.
I'm just wondering about other uses where such a system would be useful aside from asking it via Signal or WA for the weather.
I run it in a MacOS vm, fwiw, for certain tasks. I usually hand it a repo and say “install all build dependencies” and it’ll click through installation GUIs where necessary. I wonder if the TOS of an application is enforceable if a bot agreed and a human never saw it.
It can UI test web apps if that’s part of your job, but it can also just use the browser for tasks. I’ve had it research flights by checking different dates automatically and compiling the results for me. I’ve heard others use it to handle building a shopping cart for instacart/grocery delivery.
Probably the hero usecase for most people would be app/web development. It can control either a browser or a device simulator and can iteratively build and test the application, by clicking real buttons.
People need help with shopping now?
For serious stuff, is part of the shopping flow.