Abstract As demands for behaviorally sophisticated software grow, agent-based systems are increasingly being employed. Software agents are frequently applied to large, complex systems that involve interdisciplinary development teams. These complex systems have proved to be challenging to develop and evolve, even for the most competent software engineers. Taking lessons learned in other engineering disciplines such as computer and architectural engineering we investigated a model-based engineering approach called Model-Driven Architecture (MDA) to automate, whenever possible, the development and evolution of agent-based applications. In our investigation, we use the Cognitive Agent Architecture (Cougaar); one of the most mature and sophisticated collaborative agent-based architectures. MDA and Cougaar served as the primary components and implementation platform for our research. In this paper we present our approach and demonstrate how MDA is effective for producing sophisticated agent-based systems. A key challenge was found in designing a flexible meta-model framework that would accommodate both top-down domain information and bottom-up platform specific constructs, as well as the transformations and mappings between them. We employed a General Domain Application Model (GDAM) as the platform-independent model layer and General Cougaar Application Model (GCAM) layer as the platform specific model respectively. Domain-level requirements are formulated using a XML Process Definition Language (XPDL) based graphical editor and are the refined through a series of model transformations (via the underlying metamodel) to systematically generate the agent-based software system. Through an illustrative case-study, we report on the feasibility, strengths and limitations of the model-based approach as it was investigated with the Cougaar.