Retrying the connection even when the agent is not yet created in the master was something a cloud implementation could rely on (one of our proprietary implementations does).
IMHO this was a breaking change. There should be at least one way to keep the previous behaviour (with a new parameter, or maybe reusing the existent "noReconnect").