A new approach is presented for managing distributed database systems in the face of communication failures and network partitions. It offers high availability and, at the same time, guarantees some meaningful correctness properties. The approach is based on the idea of fragments and agents. It does not require prompt and correct detection of partitions and other failures.