Authors: Vjatšeslav Antipenko and Raimundas Matulevičius
Abstract: This paper presents a multi-agent architecture for formalising security risk elicitation from non-technical stakeholders. While conversational AI can support dialogue-based knowledge extraction, it lacks mechanisms to ensure structural consistency and traceability. To address this limitation, the proposed Agentic FAST Assessor integrates Work System Axioms and the FAST method into a coordinated validation process. Stakeholder statements are incrementally translated into model elements and evaluated against explicit system constraints before being incorporated into a persistent representation. This architecture separates conversational interaction from formal model construction, reducing ambiguity and improving auditability. The architecture is discussed in relation to emerging academic and industry guidelines for multi-agent systems, demonstrating how its orchestration, task separation, and validation mechanisms align with design practices. The study contributes an architectural approach for constraint-based formalisation of security risk elicitation.