Abstract
Although the notion of conversations has been discussed for some time as a way in which to provide an abstract representation of extended agent message exchange, there is still no consensus established concerning how to use these abstractions effectively. This paper describes a layered approach based on coloured Petri Nets that can be used for modelling complex, concurrent conversations among agents in a multi-agent system. The approach can be used both to define simple conversation protocols and to define more complex conversation protocols composed of a number of simpler conversations. With this method it is possible (a) to capture the concurrent characteristics of a conversation, (b) to capture the state of a conversation at runtime, and (c) to reuse conversation structures for the processing of multiple concurrent messages. A prototype implementation of such a system with some examples is described.