Thanks for the clarification @felixbarny
(Without intending to hijack this thread) Just a note on why we we used a client-side proxy: it was easier for us to abstract the resultant APM Servers this way. As the server_urls config is not dynamically reloaded, by putting the APM Servers behind HAProxy or BigIP, we can easily scale the APM Servers as well as move the target APM Servers whenever we want without having to reload the instrumented services. I could not find any other way to do this in such a scenario. We're definitely open to suggestions 
Thanks!