Your question is absolutely correct.  Standard Agents will handle queued items, till the queue has some items, let it be any no of nodes they will handle till the queue has items.  Node classification from 7.3 onwards facilitated to easily identify some of the nodes with a definite name instead of node encrypt code. Pre 7.3 environments used to have schedule data instances on each node, whereas 7.3 onwards we can dedicate particular node as "Background Processing" or "WebUser" which will be running schedule data agents of specific type on specific node only.  Whereas "Runonallnodes" is a type of node definition which is all purpose node that schedule data agent can run on any node which of this type. This is as same as pre 7.3 environments.  Incase of Advanced agents where the action is recurring or scheduled, which may need to send a report, or notification with a list of results should be a one node action which should not be interfered or repeated by any other node. Load balancing comes in the way as the other advanced agent actions can be handled by other nodes, not picking the ongoing one or sharing the same execution with some other node.