The complexity of large distributed real-time systems can be reduced by partitioning the system into a set of almost autonomous subsystems that are connected by stable control-free interfaces called temporal firewalls. A temporal firewall provides an understandable abstraction of the subsystem behind the firewall, confines the impact of most changes to the encapsulated subsystem, and limits the potential of error propagation. The paper describes the stable properties of temporal firewalls and discusses where, in a large distributed real-time architecture, temporal firewalls should be placed. The final section investigates some possibilities of how to renegotiate the static properties of temporal firewalls in case the controlled object changes it behavior substantially.