In distributed object systems, remote method invocation attempts to abstract communication between different machines to the level of object method invocation. Link failure in such systems often appears as a failed method invocation, which must be explicitly handled outside the remote object infrastructure. This usually implies that techniques to handle the effects of link failure are embedded within application functionality. The resulting application can become unnecessarily specialized to a particular environment, restricting the situations where it could be used. In this paper, we explore the possibility of cleanly decoupling policies for handling a link failure from the application.With this goal in mind, we present extensions to the Java RMI system in the context of a distributed system with mobile hosts. An intermittent communication link is typical in such systems, and applications use techniques such as data caching or queuing of method calls to alleviate the effects. We show how such techniques and policies can be implemented generically so that not only does the application become oblivious of them, but also policies can be chosen dynamically depending on a particular situation.