Generative artificial intelligence (AI) has been extensively used for cognitive tasks such as question answering. We investigate the use of generative AI for the metacognitive task of self-explanation in intelligent agents to answer Why?, What?, and How? questions about how the agent works, thereby enhancing transparency of its reasoning and increasing trust in its conclusions. We describe an architecture in which the metacognitive layer in the intelligent agent contains a model of how the base agent works. We represent this self-model in the form of a Task-Method-Knowledge (TMK) representation that captures how the methods and the knowledge of the agent accomplish its tasks. Given a question about the knowledge, reasoning or outputs of the base agent, we use generative AI to search through the TMK model of the agent, collect snippets relevant to the question, and compose an explanation as an answer. We present preliminary evaluation of this technique for self-explanation in the context of Virtual Experimental Research Assistant (VERA), an inquiry-based environment for learning by modeling and simulation.