@tcballard: An agent should not take over your desktop

Tom Ballard
@tcballard
An agent should not take over your desktop
Cua declared its Omarchy driver stable today. The useful promise is that an agent can work in a window without borrowing your mouse.
The team worked with DHH, Spencer Bull and Vaxry on a separate synthetic cursor inside Hyprland, the software that manages Omarchy’s windows and input. Hyprland routes the agent’s cursor and yours separately. You should not have to stop using the computer just because an agent needs to click something. Read Cua’s announcement and its explanation of separate input.
Today’s news is the stable announcement, not the invention of that cursor overnight. Cua Driver 0.28.3 arrived late yesterday with Hyprland input repairs and Linux ARM64 downloads. Cua says the Omarchy integration remains on Edge. Its stable driver label does not mean the integration has reached Omarchy’s stable channel.
I want an agent to do useful work beside me. Handing over the whole desktop and waiting for it to finish is a poor bargain.
Separate input does not establish safe behaviour, though. An agent can still do the wrong thing with its own cursor. The permissions beneath it matter just as much.
Permission needs an end time
Omarchy merged a set of security repairs today that makes temporary authority harder to retain or reuse.
An update can need administrator permission. That should not give every script it runs reusable access to the same permission.
The passwordless sudo repair puts an expiry deadline into the permission rule itself. Suspending the laptop should not extend a grant beyond the time the user chose. A timer removes the expired rule, and a cleanup failure remains an error.
A second repair, authored by Afonso Oliveira, keeps user-controlled update hooks and package-building steps outside authorization created by the updater. Privileged stages finish before that user code runs, and the workflow revokes its authorization on completion, cancellation or failure.
The update sleep-prevention helper also gets stricter cleanup. It checks process ownership and identity before stopping anything, addressing a route that could make cleanup terminate another account’s process.
These changes merged into Quattro development. Matching release packages and complete upgrade validation remain outstanding. They are not a new stable security release, nor evidence that every agent action is now contained.
I want the convenience of temporary permission. I also want the computer to mean it when that permission expires.
Two packages move in different directions
Yesterday’s Aurora kernel publication acquired an important restriction today. The new Apple Silicon build is Edge-only, pending M1 and M2 cold-boot qualification. Its package also stops advertising itself as generic linux, which could cause a non-Apple ARM installation to select the Apple-only kernel.
The corrected build reached Edge. The PR says the previous build still needs manual removal from the release-candidate database; the merge does not perform that cleanup.
Meanwhile, Task Manager 0.1.1 reached Edge, RC and stable for x86_64 this evening. This one is mine, so I have an obvious personal interest in seeing it become an ordinary package people can install.
The maintenance release reduces GPU sampling work and makes process snapshots specific to the page being viewed. The package request records successful builds and earlier XPS acceptance, while leaving a fresh XPS pass for the changed sampling path outstanding. NVIDIA measurements do not establish AMD or Intel coverage.
A task manager should help explain where the machine’s resources are going. It should not become another reason to ask why the machine is busy.
💬 6 · 🔁 15 · ❤️ 148 · 👁 13K
3:34 AM · Sep 26, 2026 · View on X
Replies (8)

creedants @creedants
File chooser still walks out of the window. Pointer never moves, and the keys follow whatever just took focus.

ZEK5 @0zeksss0
A meandering article though I agree, models should be in the background, but that’s a skill issue. A vast majority of models know more than I, scary for sure but the models only do what you tell them to do. Agents shouldn’t take over desktop focus but, I’m giving it its own host

Petey struggles @peteystruggles
what is the difference between "useful promise" and "unuseful promise"

Jojo @Bangeek
Great explanation !

DHH @dhh
Thrilled to announce that @ThePrimeagen has joined Omarchy Core to lead Agentic QA. His new Oligarchy harness will soon be vetting all our upcoming releases using @digitalocean droplets, which will host agents running on our @AIatMeta tokens. Full cricle!

ThePrimeagen @ThePrimeagen
Hello,
So I have been pursuing this idea of agent based automation for a while now, and now I get to put my ideas into a real project. Pretty excited!
I'll try to do a proper write up, milestones, and expectations this week.

dax @thdxr
this only works in opencode. yes. really. only in opencode. it's entirely the harness. we do harness engineering. it's very good. it's not the model. it's opencode. yes
▶

Josip @l3d1c
>be senior devops engineer in Germany
>master's degree, 10 years experience, five interview rounds
>72,000€ before tax
>C2 German required
>All team members based in Asia speaking English
>mandatory office attendance to sit on Teams with people at another mandatory office
>"we
Comments (0)
No comments yet