POMDPs are a popular framework for representing decision making problems that contain uncertainty. The high computational complexity of finding exact solutions to POMDPs has spawned a number of research projects which are investigating means of quickly finding high quality approximate solutions. This work uses information gained at runtime to reduce the cost of reasoning in POMDP type domains. Run-time dynamic decision network construction is introduced and the role it plays in an agent architecture for reasoning in POMDP domains targeted to run locally on personal digital assistants (PDAs) is evaluated. The ability to suport reactive, high quality approximate planners on lightweight devices should help to open the door for localizable and personalizable POMDP applications.