but I'd prefer to have a new instance deployed for each new experiment and that it also terminates when no new experiments are queued

I'm not objecting, just wondered on the rational behind the decision 🙂
Back to the AWS autoscaler:
Basically if you have the services-agent running on your cluster, it will just run the aws-autoscaler for you 🙂
The idea of the service-agent is to run logic/monitoring Tasks suck as the aws autoscaler. Notice that service-mode means multiple job per agent, contrary to the default one task per agent at any given time.
If you want you can package the aws-autoscaler example inside a docker and just spin it, you can use the clearml-agent docker file as a good starting point. wdyt?