We were actually previously using vagrant (-libvirt) for our CI, but moved away from it to this because vagrant does not seems to work very reliably when run in parallel. For example image downloading doesn't seem to have a filesystem lock, so one would succeed and the others failed.

In general, we wanted to be able to spin up our test VMs under docker containers to eliminate issues with differences between our CI shell-runners. Using libvirt based tooling makes that pretty hard. Calling to qemu directly, on the other hand, makes this easy.

There are some other benefits as well such as having direct console access which is not provided by vagrant, but can be very useful for kernel tests.