Noticed here and copied here: if you try to set up a Docker Compose demo of a Jenkins master plus a "JNLP" agent, the agent container will start quickly, try to connect, get a failure, and exit before the master comes up. So you need to manually bring up the master, wait a bit, then manually bring up the agent.

There is a flag to suppress automatic re-connection after losing a connection to the master. I am asking for something like the opposite: reconnect before making an initial connection.

This would be in hudson.remoting.Engine I suppose, though the entry point I was using was hudson.remoting.jnlp.Main: the first flag causes errors on the first attempt to be fatal.