How can we build systems with 'commonsense', the thinking skills that every ordinary person takes for granted? In this paper, we describe a multi-agent architecture for enabling commonsense reasoning which is in development at the Media Lab. The system reasons about the kinds of fundamental entities that show up in nearly all situations -such as people, objects, events, goals, plans and mistakes. The architecture supports multiple layers of reflective reasoning, mechanisms for coherent

more »

... across multiple representations, and large-scale control structures called 'ways to think'. We first describe the main features of our architecture and then discuss its application and evaluation to an artificial life scenario. At the Media Lab we are designing a system for 'computing commonsense', an AI operating system for the next generation of computing devices. This effort encompasses both large-scale knowledge acquisition, such as Singh's Open Mind Common Sense project [2], and Lenat's Cyc project [3] , and work on the architecture for reasoning described here. This system will be applicable to a broad spectrum of applications, from high-end servers to consumer devices such as cell phones, cars and houses. We envision applications such as: • telephones that understand the social relationships among participants, • computers that see bugs and failures in their own processing and repair and update themselves,