The area of software agents has experienced an exponential growth during the past decade, and is now being given a further boost by the introduction of global distributed computing services, such as Globus and Legion. The focus of the research has gradually shifted from single agent architectures to multi-agent systems and agent societies. In a Grid computing environment, agents should be able to discover efficiently other agents based on the computational services they offer or their characteristics (agent discovery). Existing systems either ignore this issue or use simplistic organisation models, which act as passive yellow pages thus keeping the discovery process separate from the computation. In this paper it is argued that the agent discovery can be coupled with several aspects of the computation such as access control and customisation resulting in a better sharing, use and management of the information held by the agent discovery system. A novel architecture is presented in which discovery messages and discovery paths are mutable, active entities, which interact with each other as peers making the organisation system dynamic in nature. Queries traversing the system can be reformulated while at the same time the system itself can change depending on the nature and volume of the query traffic. Furthermore it is shown that the organisation nodes of the proposed architecture can serve as re-usable components for building more complex, composite nodes from existing ones.