EONIAN: User-Guided Multi-Agent System for Cross-Provider Data Retrieval and Analysis

Data-driven scientific workflows increasingly require integration across heterogeneous repositories with divergent metadata standards and access interfaces. Existing LLM-based multi-agent systems rely on implicit conversational state for workflow coordination, making them fragile under the context pressure of long-horizon, multi-repository tasks. We present EONIAN, a multi-agent framework that decouples workflow state from conversational history through structured JSON representations: a planning blueprint and an execution state machine, which enable stateful orchestration, checkpoint-based recovery, and explicit Human-In-The-Loop intervention at critical decision boundaries. Contract-grounded planning constrains agent selection and artifact routing to deterministic structural knowledge, reducing reliance on LLM inference for task decomposition. A prototype implementation built upon iChatBio demonstrates EONIAN’s feasibility across four case studies in environmental and biodiversity science with progressive complexity, spanning single-repository visualization to cross-institutional and cross-domain semantic alignment. Token consumption analysis shows that analysis script generation and execution dominate the total cost, while orchestration overhead remains stable at 3-7% of input tokens across all cases — consistent with the system’s design rationale.

Recommended citation: Zehua Zhu, Michael Elliott, Chia-Chen Kuo, Jose A.B. Fortes, Julia A. Jones, Renato J. Figueiredo, "EONIAN: User-Guided Multi-Agent System for Cross-Provider Data Retrieval and Analysis", 22nd IEEE International Conference on e-Science, 2026 https://ieeexplore.ieee.org/xpl/conhome/1001511/all-proceedings