CowNet-AI: A Multi-Agent, Explainable Decision Support System for Social Network Analysis and Welfare Management in Dairy Cattle

Dairy farms increasingly deploy sensor networks that capture high-frequency behavioral and spatial data, yet transforming these multimodal streams into actionable welfare insights remains difficult. This work presents CowNet-AI, a modular, multi-agent conversational decision support system that performs social network analysis in near real time, generates explainable welfare recommendations, and supports simulation-based scenario testing. The system integrates a behavioral simulation engine that produces synthetic herd interactions, a data ingestion pipeline, social network analysis modules that compute centrality- and isolation-based metrics, and a LangGraph-based multi-agent architecture comprising supervisor, data loader, SNA, research, simulation, response, and report-generation agents. System components are validated using the MmCows dataset, and the agentic pipeline is demonstrated through natural-language farmer queries. Comparative experiments across two evaluation settings show that CowNet-AI achieves perfect intent routing accuracy and a low hallucination rate (5%) while consistently outperforming baseline single-LLM and static rule-based SNA systems. Results further demonstrate that CowNet-AI can autonomously route open-ended questions, compute relevant behavioral metrics, retrieve research-based context, and provide personalized, explainable recommendations, with end-to-end latency varying by agent configuration and query type. To our knowledge, this is the first fully open-source, agentic, explainable AI platform designed specifically for managing social structure and welfare in dairy cattle. Remaining challenges include mitigation of large language model hallucinations, formal quantitative evaluation against expert benchmarks, and computational scaling to larger herds and longer time horizons. Overall, CowNet-AI advances precision livestock farming by bridging the gap between sensor data collection and welfare-oriented decision making.