This paper proposes a classification framework to help with the understanding and integration of contributions in the field of adaptive multi-agent systems. This framework is used to highlight gaps in the field and derive directions for further research. The need for this framework has arisen from the proliferation of fragmented streams of research, aiming to enable adaptation of agent systems to rapidly changing circumstances and requirements. Multi-agent systems are purported to provide flexible support for users and organisations in dynamic and complex open environments because of their capabilities of autonomous problem-solving, However, exploring the boundaries of flexibility quickly uncovers limitations when agents have to adapt to situations which have not been considered during design time, This issue has been addressed by different research groups using approaches such as flexible systems, evolutionary computation, control systems, and complex adaptive systems, Nevertheless, exchange of ideas between different groups is rare, and systematic analysis of achievements is overdue. The classification framework proposed here is used for such analysis and covers both the analysis and the results in terms of directions for future work. © Springer-Verlag Berlin Heidelberg 2006.