Governing the Transition to Action: An Agentic Architecture for Situation-Aware LLMs

Abstract

This paper proposes a framework of ten architectural requirements for AI decision support in safety-critical domains. Derived from Endsley’s SA model and the distributed SA perspective, the requirements span real-time perception, structured comprehension, projection, temporal depth, transparency, operator state modelling, auditability, governed activation, inter-agent coherence, and self-monitoring. We evaluate three classes of LLM-based architecture against these requirements and demonstrate that only agentic workflows with knowledge graph grounding can satisfy them. A technical architecture for a simulated air traffic control environment demonstrates how each requirement maps to concrete infrastructure components, and identifies which requirements are readily met and which remain open research challenges.

Keywords: Situation Awareness, Artificial Situation Awareness, Agentic AI, Knowledge Graphs, Human-AI Teaming, Air Traffic Control, Large Language Models

DOI: 10.54941/ahfe1007315