What we have measured
Running agents on desktops for a living produces numbers, and most of them are not published anywhere. These pages are the useful ones: which model actually drives a Linux desktop, whether the accessibility tree beats screenshots, what happens when a person and an agent share one mouse, and what a fleet of machines costs for an eval. Where the data is somebody else's, we say so beside the number.
Scale AI's dataset, recomputed 2026-11-20
Which model drives a desktop best
Pass rates for eight models driving a Linux desktop, from Scale AI's per-trial Cua Driver Bench v2 dataset recomputed on 2026-11-20, with the cost a trial and what changes when a person is using the machine.
Our own lab, 2026-11-15
Screenshots or the accessibility tree
We ran one real desktop task ten times through Kubi's MCP server, from screenshots only, from the accessibility tree only, and from both. Times, turns, tokens, and where the turns actually went.
A persistent screen keeps its disk
Browser logins an agent keeps
How to let an agent reuse a signed-in browser: one profile on a desktop that stays on, a person who types the password, and Playwright attached to the browser that is already open.
No idle shutdown
An agent that keeps running when you close your laptop
How to keep an agent running with your laptop shut: install it on a cloud Linux desktop, run it as a service, keep its key as a secret, and what the month costs.
409 CONTROL_HELD while you hold it
Watching an agent, and taking the screen back
How a person shares one desktop with an agent without the two fighting over the mouse: a live view, a lock with a clear error, an idle release, and a view-only link.
Disposable screens, by the minute
A benchmark on a hundred screens
What it takes and what it costs to run an eval across many cloud Linux desktops: templates, the API that creates and deletes them, per-minute billing, and the arithmetic.










