ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Aleks 2.0: Enhanced Single Agent MCP System for Autonomous Research in Plant Science

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600898.(doi:10.13031/aim.202600898)
Authors:   Daoyuan Jin, Ruiming Du, Shivranjani Baruah, Adam Bogdanove, Kaitlin M. Gold, Yu Jiang
Keywords:   LLM; AI agent; Scientific discovery; Knowledge graph; Model Context Protocol

Abstract. AI agent systems powered by large language models (LLMs) have demonstrated significant potential for automating scientific discovery workflows. Our previous work introduced Aleks, an AI-powered multi-agent system that autonomously conducts data-driven research in plant sciences by integrating domain knowledge, data analysis, and machine learning. While Aleks achieved full autonomy in addressing research questions, its capabilities were constrained by simplified knowledge organization and a limited toolset. Here we present Aleks 2.0, an enhanced system featuring three major improvements. First, a LightRAG knowledge graph replaces flat summarization as the retrieval backend for the Domain Scientist, enabling structured entity-level retrieval and cross-paper reasoning from scientific literature. Second, a template-based code generation strategy stabilizes machine learning pipeline construction under locally deployed small-parameter language models by externalizing factual API knowledge and structural constraints into a reusable scaffold. Third, a Model Context Protocol (MCP) based architecture provides a standardized, pluggable interface through which domain-specific tools contributed by external collaborators can be registered and invoked directly within the agent pipeline without modification to the core system. Results demonstrate that graph-based retrieval outperforms flat summarization across all evaluated question categories, template-based instruction achieves reliable code generation where freeform conditions fail, and the full pipeline operates end-to-end without human intervention on a rice bacterial blight severity prediction task.

(Download PDF)    (Export to EndNotes)