CeresAgents | Safety-Aware Crop Disease Management

CeresAgents (A multi-agent diagnosis-to-prescription framework for safety-aware crop disease management)

Lufu Qin, Jiaying Wang, Xiaodan Wang, Wei Guo, Yuan Li, Zaiwen Feng, Meili Wang, Juntao Xiong, Xiaohong Pan, Yong Wang, Feng Zhou, Warren Jin, Congcong Sun, Ji Zhou, and Qi Wang.

Crop disease management requires more than symptom recognition: diagnoses must account for field context, and chemical recommendations must satisfy registration and label-use constraints. Although vision models and large language models support parts of this process, coordinating diagnostic evidence and treatment decisions remains challenging. We present CeresAgents, a task-adaptive multi-agent framework for evidence-grounded and safety-aware crop disease management.

An intent-based coordinator selectively activates pathology, physiology, and chemistry specialists according to consultation requirements. Through a shared blackboard, specialists exchange diagnostic findings and contextual evidence, enabling structured reassessment when pathological and physiological interpretations conflict. Visual disease models and hybrid-contextual literature retrieval support disease identification and differential reasoning, while a chemistry specialist uses a pesticide knowledge graph to verify candidate recommendations against available crop–disease registration and label-use information.

We evaluate CeresAgents on CeresBench, a manually curated benchmark of 267 cases spanning seven task tracks. CeresAgents achieved 94.89% diagnostic accuracy and an 84.60 end-to-end quality score, outperforming four direct foundation-model baselines. Compared with a single agent using the same backbone and evidence tools, CeresAgents improved safety by 16.69 points.

Contributions.

We develop a task-adaptive multi-agent architecture that coordinates pathology, physiology, and chemistry specialists through a shared blackboard.
We integrate visual analysis, hybrid-contextual literature retrieval, and pesticide-registration knowledge to support evidence-grounded recommendations and withhold chemical prescriptions when evidence is insufficient.
We introduce CeresBench to evaluate routing, tool selection, diagnosis, prescription, safety, conflict handling, and end-to-end advisory quality.

**Figure 1:** CeresAgents architecture, showing task-adaptive routing, specialist collaboration through a shared blackboard, evidence tools, and final response synthesis.

**Figure 2:** Task-adaptive agent activation across representative crop-health consultation situations.

**Figure 3:** Shared-blackboard coordination, showing specialist state updates, evidence tracking, and state compaction before final synthesis.

October 20, 2026. CeresAgents project page update.