Desktop tests are device tests
QA Wolf's new Windows test system treats each desktop as a claimed device. Test code stays in a Linux runner. A controller assigns one ready Windows VM. CDP reaches browsers, Appium reaches native controls, and WebRTC carries the screen back for live viewing and recording.
The useful unit is not "a Windows server." It is one desktop, claimed by one run, with one release path.
The company reports that a cold Windows VM takes two to three minutes to boot. Its warm pools hide that delay. Each run gets a VM that is already booted and logged in. After release, that VM is destroyed and the pool replaces it. A heartbeat lets the controller reclaim a claim after 60 seconds of silence.
Clean and observable are different
Single use blocks one test from inheriting another test's profile, registry keys, files, logins, and stray processes. It does not create browser-style evidence by itself. QA Wolf added a desktop agent, screen recording, diagnostics, and logs because a remote desktop starts as a black box.
That split should survive in your own test card. Image identity answers what booted. The artifact digest answers what arrived. The recording and action log answer what the runner saw and did. The assertion answers what passed. The release record answers what the controller destroyed.
Disposable has sharp edges
Microsoft's Windows Sandbox makes the boundary concrete. Closing the sandbox deletes its software, files, and state. A restart inside the sandbox can preserve data, though. Networking is enabled by default, and Microsoft warns that this can expose an internal network to untrusted applications.
QEMU offers another mechanism. Its -snapshot mode treats disk images as read-only and sends writes to a temporary file. A separate commit operation can write changes back. That means "temporary" is a policy plus a command path, not a magic property of the word snapshot.
For a desktop test fleet, define which writes die with the guest and which evidence leaves it. Keep recordings and logs in a named result store. Keep test credentials short-lived. Restrict artifact delivery. Prove the claim was released. A destroyed VM does not prove every copy outside that VM was deleted.