MindCare: A Hybrid Multi-Agent Architecture for Stress-Aware Conversational Support

Abstract

We present MindCare, a hybrid multi-agent architecture for stress-aware conversational support. The system decomposes a conversational mental-health support pipeline into three specialized agents: an Input Agent for interpretable stress classification, a Judge Agent for LLM-based contextual reasoning, and an Output Agent for response generation. Unlike monolithic LLM systems, this decomposition separates structured prediction from semantic reasoning and natural-language generation, allowing each component to be audited and optimized independently. To support heterogeneous inputs, behavioral and physiological features are converted into structured natural-language descriptions and processed alongside textual data. Preliminary experiments on three public stress-related datasets show that lightweight classical classifiers substantially outperform zero-shot LLMs for stress classification, with linear classifiers achieving macro F1 scores of 0.86 compared with 0.23 for zero-shot Gemini models. A pilot user study with 18 participants showed generally positive satisfaction, and classifier-conditioned routing produced a measurably different distribution of conversational responses. These preliminary results suggest that combining interpretable classifiers with LLM reasoning may provide a promising foundation for stress-aware affective conversational systems.