We propose a system architecture that facilitates dynamic assurance of autonomous systems embedding machine learning-based components. We also introduce dynamic assurance cases as a generic framework to provide justified confidence in their capabilities.