This paper proposes a unified methodology for semantic system analysis that integrates Retrieval-Augmented Generation (RAG), semantic digital twins, Agentic AI, and bidirectional semantic graph construction into a common analytical framework. Unlike conventional large language model (LLM) applications that directly generate answers from prompts, the proposed methodology first constructs a semantic digital twin of the problem domain from dynamically retrieved evidence and then performs system analysis over this intermediate model. The methodology combines forward semantic expansion from the current state with backward semantic expansion from the target or forecast state, producing an integrated bidirectional semantic digital twin that serves as the basis for scenario extraction, multicriteria evaluation, ranking, and explainable recommendation generation. A fundamental contribution of the proposed approach is the unification of decision support and forecasting within a single methodological framework. In decision-support problems, the target state is explicitly defined by the user, whereas in forecasting the target state is initially unknown and is represented by a set of alternative future hypotheses automatically generated by the LLM. Each hypothesis is transformed into its own bidirectional semantic digital twin, after which the same procedures of scenario extraction, evaluation, ranking, and explanation are applied. Consequently, decision support becomes a special case of the more general forecasting methodology corresponding to a single target state. The methodology is implemented through an agent-based execution framework in which specialized agents cooperatively construct, refine, analyze, and interpret the semantic digital twin using iterative interactions with Retrieval-Augmented Generation. Experimental case studies demonstrate that the proposed approach provides interpretable, evidence-grounded, and reproducible solutions for both decision support and forecasting while preserving a common system-analysis workflow. The proposed methodology defines a new class of Agentic Semantic System Analysis systems, in which large language models function primarily as instruments for constructing and analyzing semantic digital twins rather than merely generating textual responses. Keywords: System Analysis, Semantic Digital Twin, Agentic AI, Retrieval-Augmented Generation, Decision Support, Forecasting, Semantic Networks, Explainable AI