I had this working with older versions of Hudson, it broke on upgrade.

I can manually launch the agent, and it is configured as such.

Launch method = ssh
Availability = Take node off line when idle

then when the agent goes off line Hudson fails to restart it. I then have to
manually click the "launch slave agent" to restart it. Once this is done, the
agent runs until the agent is idle, when it will go off line again, and will
never come on line again until I click the button "launch slave agent".

The agents are running on Scientific Linux (Red Hat binary compatible OS) 32 and
64 bit, versions 4 and 5.

The following is a typical log file

[06/16/09 18:43:32] [SSH] Connection closed.

<lots of jibberish characters>

channel stopped
[06/16/09 18:43:32] slave agent was terminated
hudson.remoting.Channel$OrderlyShutdown
at hudson.remoting.Channel$CloseCommand.execute(Channel.java:623)
at hudson.remoting.Channel$ReaderThread.run(Channel.java:758)
Caused by: Command close created at
at hudson.remoting.Command.<init>(Command.java:47)
at hudson.remoting.Channel$CloseCommand.<init>(Channel.java:619)
at hudson.remoting.Channel$CloseCommand.<init>(Channel.java:619)
at hudson.remoting.Channel.close(Channel.java:663)
at hudson.remoting.Channel$CloseCommand.execute(Channel.java:622)
... 1 more

This is 100% repeatable, we are running Hudson ver. 1.310 and this bug has been
around for quiet some releases, maybe 3-4 months.