This has happened a couple of times on ci.j.io. For whatever reason, we're not entirely sure, the init script for our Windows VMs sometimes fails.
As a result, the agents are provisioned in Azure, but Jenkins fails to launch the JNLP agent on the machines. Unfortunately, Jenkins doesn't clean up these resources properly using our 20 minute idle retention policy.
I discovered this morning that we had 50 VMs online for quite a few days, which never executed a single workload because their init script failed to run, and they were not cleaned up properly by the plugin.