SAGE: A Theory-Informed LLM-Based Multi-Agent Recommendation System for Grounded AI Solutions Across Domains

Jiawei Tong, John Moraros, Shuihua Wang, Guangyu Wang, Yan Wang, Ruoxi Liao

Keywords: LLM-based Multi-Agent Systems, Responsible AI, Agent Coordination, Domain Knowledge Integration, Theory Grounding, Scalable Agentic Systems

TL;DR: SAGE addresses the Multi-Agent Responsibility Gap through five specialized LLM agents that coordinate to integrate domain theories with AI methods, achieving responsible, reliable, and scalable agentic systems for knowledge-intensive domains.

Abstract: Large Language Model-based multi-agent systems (LaMAS) show promise for solving complex problems, yet lack systematic mechanisms for responsible integration of established domain knowledge—a critical gap that risks producing technically sophisticated but contextually inappropriate solutions. To address this fundamental challenge, we present SAGE (Specialized Agent Group for Expert-grounded recommendations), a novel multi-agent framework that addresses this challenge through five specialized agents coordinating via asynchronous protocol to integrate domain theories, AI methods, and data sources. Our architecture implements three key innovations: (1) theoretical grounding mechanisms} that prioritize established domain wisdom over algorithmic capabilities, (2) coordinated specialization} where agents handle distinct reasoning tasks—scenario analysis, theory retrieval, algorithm matching, data selection, and validation—rather than monolithic processing, and (3) robustness validation through Monte Carlo simulation (N=1000 runs) with human-in-the-loop oversight ensuring R > 0.85 stability across perturbed scenarios. We validate SAGE's effectiveness in urban planning and safety, a complex domain requiring integration of social, environmental, and spatial theories. Empirical analysis of 1,123 AI applications reveals that only 1.16% demonstrate explicit theoretical integration, with 47.3% following technology-driven rather than problem-driven approaches. Through three representative case studies, we demonstrate how SAGE transforms narrow technical solutions into comprehensive, theory-grounded interventions that generate value beyond algorithmic performance metrics. Our framework establishes generalizable design principles for responsible agentic systems—agent specialization strategies, coordination protocols for knowledge integration, and validation mechanisms—with demonstrated potential for adaptation to other knowledge-intensive domains requiring systematic integration of theoretical frameworks with computational approaches.

Submission Number: 6