# What we have measured

*Kubi, cloud desktops for AI agents. https://trykubi.com/guides, last changed 2026-11-20. The page as Markdown; the same facts as the page itself.*

Running agents on desktops for a living produces numbers, and most of them are not published anywhere. These are the useful ones. Where the data is somebody else's, the page says so beside the number.

- **[Which model drives a desktop best](https://trykubi.com/guides/which-model-drives-a-desktop-best)** (Scale AI's dataset, recomputed 2026-11-20, updated 2026-11-20). 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.
- **[Screenshots or the accessibility tree](https://trykubi.com/guides/screenshots-or-the-accessibility-tree)** (Our own lab, 2026-11-15, updated 2026-11-15). 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.
- **[Browser logins an agent keeps](https://trykubi.com/guides/browser-logins-an-agent-keeps)** (A persistent screen keeps its disk, updated 2026-11-23). 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.
- **[An agent that keeps running when you close your laptop](https://trykubi.com/guides/run-an-agent-that-never-stops)** (No idle shutdown, updated 2026-11-23). 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.
- **[Watching an agent, and taking the screen back](https://trykubi.com/guides/watch-an-agent-and-take-over)** (409 CONTROL_HELD while you hold it, updated 2026-11-23). 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.
- **[A benchmark on a hundred screens](https://trykubi.com/guides/a-benchmark-on-many-screens)** (Disposable screens, by the minute, updated 2026-11-23). 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.
