In my previous post I assumed that I would need to explicitly load my child objects through my service and provide a finer grained service interface to support that model.  After digging through Udi Dahan's blog posts again I found a very relevant post about using a fetching strategy to specify how deep you want the object model to be for any specific request.  By being able to specify a fetching strategy in the consumer code I can return the proper size object graph without making my service interface chatty.

This does leak some of the service logic into the client code, but really I'm OK with that.  The alternative is to explicitly call each service method which I think is harder to maintain and makes the object model pretty difficult to use. 

After using Mule ESB at my last job I've found it kind of hard to go to plain old web services.  I really like the messaging paradigm and separation of concerns an ESB provides.  Now if I could just find the SVN repository for NServiceBus.