The ternary alphabet representing classifier conditions in traditional XCSs suffers from a lack of richness and flexibility and therefore limits the capability of the algorithm to cope with complex and large-scale problems. Code Fragments (CFs), a form of tree-based programs, were introduced into XCS to represent classifier conditions and actions. CFs are a rich and flexible representation, which can theoretically encode any function. However, its flexibility also comes with adversary impacts. Phenotypically same CFs can have many different representations, but the comparison of any two CFs is done syntactically. This leads to the undesirably large search spaces and slow learning performances. In addition, the existing methods to generate CFs and transfer CFs between problem domains are only guided from the current instance. This mechanism is not likely to create generalised solutions which can match large numbers of instances. This research will develop novel methods to overcome these limitations and improve the quality of the transferring of knowledge. This is done by addressing the problems of problem decomposition and finding appropriate functions to decomposed subproblems. This improvement is expected to allow building an efficient multi-agent system using CF-based XCSs with the ability to learn problems concurrently and decide automatically which knowledge is useful to an agent of the system. Thus, the agents of the system can support each other efficiently to learn large-scale and highly complex problems concurrently. The implementation of this multi-agent system is expected to be a step closer to a Cognitive System.