This paper describes the development of an Autonomous Underwater Vehicle (AUV) and its intelligent control system. The architecture of the control system is reactive, rather than attempting to model the entire underwater environment, the robot instead relies on a set of "reflexes" coupled with a number of higher-level task planning modules. The "reflexes" endow the robot with failsafe survivability; they ensure that it will avoid potentially dangerous situations, thus freeing the task planning modules to operate at a high level of abstraction without having to worry that they may have missed some potentially hazardous feature of the underwater environment. The control architecture used by the robot is based on the Subsumption architecture.