CowNet-AI: A Multi-Agent Decision Support Framework for Social Network–Driven Welfare Insights in Dairy Cattle

Dairy farms increasingly deploy sensor networks that generate high-frequency behavioral and spatial data, yet translating these multimodal streams into actionable welfare insights remains a key challenge. This study presents CowNet-AI, a modular, multi-agent conversational decision-support framework that integrates social network analysis with large language model–driven agents to deliver explainable welfare risk indicators and simulation-based insights. The system comprises a behavioral simulation engine for synthetic herd interaction generation, a data ingestion pipeline, social network analysis modules computing centrality- and isolation-based metrics, and a LangGraph-based multi-agent architecture including supervisor, data loader, SNA, research, simulation, response, and reporting agents. The framework is evaluated using the MmCows dataset, with demonstrations conducted through natural-language farmer queries. Experimental results show that CowNet-AI achieves high intent routing accuracy and a low hallucination rate (5%), outperforming baseline single-agent LLM and static rule-based SNA approaches. The system effectively integrates behavioral metrics with contextual knowledge to generate personalized, explainable recommendations, while enabling autonomous query routing and scenario-based reasoning. End-to-end latency varies depending on agent configuration and query complexity. This work represents a pilot-scale validation conducted on a limited dataset, establishing the feasibility of agentic, explainable AI systems for livestock welfare intelligence. Current limitations include reliance on behavioral proxies without direct clinical validation, dependence on LLM-based evaluation, and computational scaling challenges for larger herds. Future work will focus on multimodal data integration, expert-in-the-loop validation, and system optimization for real-time deployment.