We discuss motivation and design ideas for a formal semantic framework with tool support for specifying, analysing and reasoning about secure agents and secure agent architectures. Within this framework we will be able to define executable models of agent systems, hostile environments, and mechanisms for control, detection, and protection. These models will serve as a basis for search and model-checking analyses, definition of abstraction mappings, and rigorous proofs of general properties. We have in mind not only the standard security concerns of authentication, authorization, integrity, and confidentiality considered in this more complex setting, but also concerns such as privacy, stealth, and information flow and agregation. An objective of the secure agent framework should facilitate identification of vulnerabilities and assumptions at different levels of abstraction of system design.