MUM is a knowledge-based consultation system designed to manage the uncertainty inherent in medical diagnosis. The primary task of the system is to plan which questions, tests, and treatments to order at each point in a consultation given its current uncertainty knowledge about the patient's diseases. Managing uncertainty means planning what to do when uncertain; it is suggested that this ability must be designed in, not added on, to the architectures of knowledge-based systems. MUM is based on one such architecture, implemented as a generalized inference network and planner. The network facilitates local combination of evidence; the planner reads the state of the network after each piece of evidence is integrated and then decides which evidence to seek, based on its several goals.