This report describes the architecture of the ACLT system, which implements the ACLT coordination model completed with distributed (Prolog) agent activation. First, the fundamental requirements are specified, in order to show which abstractions have to be built in the system to support the ACLT model. The conceptual architecture of the ACLT system is then presented, with special regard to the client/server relationship between agents and tuple spaces and to the innovative server architecture which reinterprets the Linda scheme and extends it so as to support tuple space programmability through reactions. Finally, details are given with respect to the current implementation.