AgriAgent: A Multi-Agent Agentic AI System for End-to-End Fruit Disease Detection, Edibility Evaluation, and Treatment Recommendation from Leaf and Fruit Imagery

Sushant Patil¹, Mahesh Tupe², Pranav Udmale³, Siddharth Ekatpure⁴, Prof. Sarika Jadhav⁵,

 Prof. Suvarna Bahir⁶

¹Department of Computer Science and Engineering, [PVPIT]

²Department of Computer Science and Engineering, [PVPIT]

³Department of Computer Science and Engineering, [PVPIT]

⁴Department of Computer Science and Engineering, [PVPIT]

⁵Department of Computer Science and Engineering, [PVPIT]

⁶Department of Computer Science and Engineering, [PVPIT]

Abstract — Large Language Models (LLMs) have demonstrated strong potential for agricultural decision support, yet most existing systems address disease detection and treatment recommendation as separate, disconnected tasks, and few provide a single, verifiable pipeline from raw fruit imagery to an actionable agronomic decision. This paper presents AgriAgent, a multi-agent agentic AI system that accepts a fruit or leaf image as its sole input and produces a complete output comprising disease identification, fruit edibility classification, and specific pesticide and fertilizer recommendations. The system is organized as a pipeline of specialized, independently invokable agents — a Vision Detection Agent, an Edibility Reasoning Agent, a Treatment Recommendation Agent, and an Orchestrator Agent that coordinates control flow, retrieval, and confidence-based escalation between them — rather than as a single end-to-end model or a monolithic prompt. A domain knowledge base constructed from agricultural extension guidelines and pesticide/fertilizer dosage references constrains the Treatment Agent’s output to safe, label-consistent recommendations, in the spirit of domain-knowledge-constrained retrieval-augmented generation. We evaluate AgriAgent on a curated multi-class fruit disease image dataset against three baselines: a zero-shot multimodal LLM (GPT-4o), a fine-tuned CNN classifier (ResNet-50), and a single-stage RAG pipeline without agentic orchestration. AgriAgent achieves a disease classification accuracy of [X]%, an edibility classification accuracy of [X]%, and a treatment-recommendation validity rate of [X]% as judged against extension-guideline ground truth, outperforming the zero-shot LLM baseline by [X] percentage points and the single-stage RAG baseline by [X] percentage points, while reducing unsafe or out-of-range pesticide dosage recommendations by [X]% relative to an unconstrained LLM baseline. These results indicate that decomposing the fruit-disease-to-treatment pipeline into specialized, orchestrated agents with an explicit domain knowledge constraint layer yields more accurate and safer agricultural recommendations than either a single large multimodal model or a non-agentic RAG pipeline.

Index Terms— Agentic AI, Multi-Agent Systems, Large Language Models, Retrieval-Augmented Generation, Fruit Disease Detection, Computer Vision, Precision Agriculture, Pesticide Recommendation, YOLOv8, Domain Knowledge Constraint.