pipe.start_locally() will run the DAG compute part on the same machine, where pipe.start() will start it on a remote worker (if it is not already running on a remote worker)
basically "pipe.start()" executed via an agent, will start the compute (no overhead)
does that help?