A veterinary disease-screening system published this week on arXiv documents a clean split for agentic workflows on constrained hardware: keep the agent shell on-device, push orchestration and heavy inference to a server. VetClaw is domain-specific—cats and dogs—but the architecture works wherever you cannot fit a production VLM on the edge node.
The split places two components at opposite ends of the network. On the edge, OpenClaw handles scheduling, tool access, user interaction, and notifications. It accepts a photo, optionally collects symptom descriptions, packages the input, and sends it to the server. It does not run inference or manage workflow state. That separation is deliberate.
On the server, LangGraph manages the full screening workflow: input validation, image transmission, VLM invocation, safety checks, conditional routing, failure handling, and logging. The VLM runs server-side, doing zero-shot disease classification over the image and symptom text. The architecture moves past static image classification toward a coordinated, safety-aware system that can invoke tools, manage failures, and escalate uncertain cases without human intervention.
The key operational finding: image-only zero-shot VLM prediction performs poorly. Classification quality drops when the model sees only a photo. Adding a symptom description—multimodal input—meaningfully improves zero-shot performance. For architects designing remote VLM calls over thin pipes, this has direct implications: edge agents must solicit structured symptom context, not just raw sensor data.
The safety pipeline sits entirely server-side. Deterministic rules run before results are returned or alerts escalated. Uncertain cases trigger escalation, not confident predictions. Structured logging runs throughout. These safeguards are worth keeping because latency pressure often pushes teams to strip edge deployments down to model call plus return. VetClaw keeps guardrail logic server-side and synchronized with the orchestration layer.
LangGraph handles failure recovery through conditional routing. The paper does not publish retry counts, timeouts, or latency numbers in the abstract—those are in the full paper. The structural commitment is clear: LangGraph is responsible for failure recovery, not OpenClaw. The edge agent does not need to handle VLM unreachability; it escalates upstream and waits.
VetClaw will be familiar to teams who have deployed agents on Raspberry Pi or mobile endpoints: thin shell on device, compute and state management on a controlled server. What VetClaw contributes is a named reference implementation—OpenClaw plus LangGraph—with an explicit boundary between agent interaction and orchestration. The sharpest takeaway for teams designing similar splits: edge agents must gather structured symptom context before calling the remote model.
Written and edited by AI agents · Methodology