SOS (SOMIW Operating System) is the result of a four-year effort at INRIA to define an object-oriented operating system. SOS provides support for arbitrary, user-defrned, typed objects. The system implements object migration; this mechanism is generic, but can be tailored to specific object semantics thanks to the prerequisite and upcall concepts. SOS also supports Fragmented Objects (FOs), i.e. objects the representation of which spreads across multiple address spaces. Fragments of a single FO are objects that enjoy mutual communication privileges. A fragment acts as a proxy, i.e. a local interface to the FO. All the other mechanisms of SOS are built upon these basic abstractions. Thanks to prerequisites, migration of data may cause the migration and dynamic type-checking and linking of the corresponding code. A distributed object manager, an object storage service, a naming service, as well as a protocol toolbox and some applications, have been built as FOs. This paper gives a detailed account of the architecture and design decisions of the SOS prototype on UNIX. rùy'e examine both good decisions and problems. The basic good decision is our simple object model, and its ability to map user-defrned semantics (policy decisions) on system-implemented mechanisms. The most important problem is the dynamic nature of Fragmented Objects, and inadequate support for them.