Mobile agents have emerged as a new paradigm for computing and are proving to be more flexible and dynamic than the average client in the client-server model. In the mobile agent model the program is sent to, and runs on, the remote machine, thereby operating closer to and more interactively with the remote document resource. While for the most part industry is currently cautious, research laboratories have embraced the concept and effective prototypes have emerged. AgentSys is one such mobile agent system developed by the Agent Group at the Multimedia Information Research Laboratory [26]. This prototype implements the protocols (AgenTask, AgenTransact, AgenTransfer) necessary to allow agents to autonomously roam around a network of digital resources, collect information intermittently, and return to the user with results. While both network bandwidth and user search time can be reduced using the mobile agent paradigm, such a system requires protocols at the agent language, ontology, transfer, and application levels. This paper proposes: (i) an agent-stack to enable a mobile agent system, (ii) the agent protocols necessary for agent-digital document interaction, (iii) a profound new way of thinking of documents within the mobile agent paradigm, and (iv) solutions to critical issues in mobile agent transfer and task specification as they relate to autonomous digital document collection. This includes both the transfer of multimedia mobile agent cargo and novel techniques for the specification of agent behavior within the document task. A brief quantitative analysis of the Web surfing paradigm versus the mobile agent paradigm is included to justify claims of network traffic savings using agents.