VetClaw: An Edge-Cloud Multimodal Agentic System for Veterinary Disease Screening

Research Question and Approach

Veterinary disease screening is challenging because animals cannot communicate symptoms, and disease presentation varies significantly across species. Existing AI systems often rely on static image classification, which lacks the contextual reasoning required for accurate diagnosis. The authors introduce VetClaw, an agentic system designed to bridge this gap by combining edge-based sensing with cloud-hosted vision-language models (VLMs).

VetClaw utilizes a layered architecture: an edge device (Raspberry Pi) runs OpenClaw to manage hardware interaction and scheduling, while a stateful workflow engine (LangGraph) orchestrates the screening process. This workflow includes image validation, symptom integration, VLM invocation, and deterministic safety checks. By separating agent interaction from workflow orchestration, the system ensures that VLM outputs are treated as diagnostic-support alerts rather than final clinical diagnoses.

The researchers evaluated two state-of-the-art VLMs (Qwen3-VL-32B and InternVL3-38B) across two public veterinary datasets. The results demonstrate that image-only classification is insufficient for reliable screening. In contrast, incorporating textual symptom descriptions alongside visual data significantly improves performance. For instance, in the Dogs Skin Disease dataset, the macro-F1 score increased from 0.31 (image-only) to 0.69 (text-image). The study highlights that while VLMs provide powerful reasoning capabilities, their performance is highly dependent on the quality of multimodal input and the implementation of rule-based safety layers to handle uncertainty and potential model errors.

VetClaw demonstrates how agentic AI can transform static prediction models into robust, safety-aware systems suitable for real-world veterinary applications. By automating the coordination between edge devices and cloud-based reasoning, the system provides a framework for early disease detection that is more reliable than traditional computer vision approaches. This architecture is particularly relevant for precision livestock farming, where early, automated screening can significantly improve animal biosecurity and health outcomes.