P-16

Jilong Liu, Hengshen Liu, Zhaoqiang Li, Jianwen Zhou, Yixue Li
Guangzhou National Laboratory

While Large Language Model (LLM)-driven scientific agents exhibit remarkable cognitive capabilities, their physical execution is severely bottlenecked by the "Sandbox Dilemma." Operating in unconstrained local environments, agents face an NP-hard action space explosion, struggle with massive multi-omics datasets, and fail to maintain reproducible provenance. To bridge this gap, we present Bio-OS NaviGen, a framework that grounds autonomous agents within Bio-OS, the real-world implementation of GA4GH Cloud Standards in China. NaviGen establishes a hierarchical architecture to decouple abstract cognitive reasoning from deterministic execution. Crucially, to accommodate the diverse landscape of agent frameworks, NaviGen pioneers a dual-mode integration strategy: it provides a standard Model Context Protocol (MCP) server for modern MCP-compatible agents, alongside a native Plugin/CLI mechanism specifically designed to empower non-MCP agents by exposing platform capabilities as modular, direct-call toolsets. Regardless of the connection mode, NaviGen injects strong domain priors by utilizing Semantic Web-based Workspace RO-Crates as machine-readable world models aligned with GA4GH standards (WES, TES, DRS). This elegantly compresses the agent's unbounded search space into structured inference while mathematically guaranteeing FAIR principles. To operationalize this framework, we introduce Bio-OS Claw, a deployable agent entity built upon the OpenClaw architecture that leverages the CLI plugin mode to bypass rigid protocol dependencies. It inherently possesses advanced cognitive skills, including automated data fetching, WDL pipeline composition, and bidirectional literature-to-workspace conversion. Building upon this foundation, we developed EpiClaw, a domain-specific capability layer tailored for epidemiology. Incorporating 57 standardized skills across multi-pathogen surveillance and predictive modeling, EpiClaw orchestrates real-world public health workflows by interacting directly with analytical cloud workspaces. Ultimately, the Bio-OS NaviGen ecosystem demonstrates how anchoring probabilistic AI reasoning within a deterministic, GA4GH-compliant infrastructure enables verifiable, scalable, and truly autonomous biomedical discovery.