About AgriSkill-Agent

Why AgriSkill-Agent?

Recognition-focused models stop at a disease label. Knowledge-enhanced assistants may retrieve relevant facts without specifying a reproducible sequence from observation to management, while general tool-using agents do not inherently preserve mask-derived measurements or pesticide-record provenance.

AgriSkill-Agent addresses this gap with explicit and testable procedures. Its central LLM activates task-specific skills while specialist models and deterministic functions retain ownership of visual evidence and measurements. Unstructured scientific literature remains distinct from structured pesticide-use data.

Comparison between a direct multimodal model and AgriSkill-Agent
Figure 1: Qualitative comparison of a direct multimodal baseline and AgriSkill-Agent on the same outdoor potato late-blight consultation.
Results and Practical Value

On the internal 140-case benchmark, AgriSkill-Agent reached 92.9% diagnosis accuracy and an 89.9 overall-quality score. The system combines phenotype observations, deterministic lesion measurement, diagnostic reasoning, scientific evidence, management options, and safety notes in one structured report rather than returning only a class label.

A blinded audit of 20 reports by three domain experts supported the automated rubric evaluation within the audited subset. The framework is intended as decision support; management recommendations must still be checked against current official registration, product labels, local conditions, and professional judgment.

Future Perspectives

Future work should evaluate larger external field datasets, add healthy leaves and abiotic stresses, and incorporate temporal and environmental context. A richer knowledge graph could represent pathogens, symptoms, resistance mechanisms, weather, growth stages, and non-chemical interventions. Model calibration and human escalation should be evaluated before operational crop-protection use.

Recorded AgriSkill-Agent workflow
Figure 2: Recorded execution trajectory for an outdoor apple-rust consultation.
Research Team

Lufu Qin, Zhengbo Liu, Peijia Yu, and Qi Wang.

Guizhou Provincial Laboratory of Big Data, College of Computer Science and Technology, Guizhou University; School of Integrated Circuit Science and Engineering, University of Electronic Science and Technology of China.

Correspondence: jgs@vip.sina.com (Z.L.) · qiwang@gzu.edu.cn (Q.W.)

Supplementary note. Further field testing and larger expert audits are warranted before operational deployment.