James E. Hanson and Jeffrey O. Kephart, IBM Thomas J. Watson Research Center
Multi-agent systems in which agents can respond to messages by automatically generating and multi-casting other messages are inherently vulnerable to a phenomenon we call a maelstrom. We define a maelstrom to be a self-sustaining chain reaction in which a single message can unintentionally trigger the generation of a rapidly growing, potentially infinite number of messages, quickly incapacitating the communications network. There is reason to fear that modest advances in agent technology and usability could lead to spontaneous maelstroms on the Internet in the near future, particularly in the realm of electronic mail. In this article we describe various classes of maelstrom that may arise due to automated forwarding of messages, and propose a novel and practical means of combatting them.