In shared-control systems, such as intelligent service robots, a human operator and an automated technical system are interdependently in charge of control. Effective shared control requires complex system architectures that provide safety and robustness in operation, while providing a user-friendly interface--provision of these dual requirements is a non-trivial task. This paper reports on an approach to addressing these issues. A dialogue centric cognitive control architecture is presented, which utilizes both agent-oriented programming and formal methods. The SharC Cognitive Control Architecture is a hybrid design that distributes control of a robot amongst a community of deliberative intentional agents. Since safety is of paramount importance in shared-control systems, high-level safety issues must be addressed within such architectures. To this end, the authors also describe a formally modeled dialogue manager that sits at the heart of the control system. The use of these distinct software paradigms is illustrated, by example, with the demonstration platform: Rolland, the Bremen autonomous wheelchair.