Distributed programming suffers from the lack of abstractions and tools required to handle and analyse the large amount of information characterising distributed systems. On the other hand, the separation of computation and coordination models definitely simplifies the design of a programming environment for distributed applications. Starting from this consideration, the ACLT coordination model extends the basic Linda kernel, by providing support for heterogeneous multi-agent systems, as well as for hybrid agent architectures integrating deduction and reaction. The design of the architectural support for the ACLT model led to the definition of a general-purpose scheme which is powerful enough to be used both for the system extension of the basic communication kernel and for building application-defined development tools. Such an approach is based on the idea of reactive communication abstractions, which can be programmed by agents according to a specification language which is rooted in the same model as the coordination language.