Title: Moving Past Prompt-Driven Scheduling: A Graph-Based Biomedical Agent Framework Native to MCP

Abstract:

While biomedical agents hold the potential to streamline intricate biological processes, existing frameworks are hindered by two primary constraints. First, bioinformatics utilities exhibit significant diversity in their interfaces and execution environments. Second, agent scheduling continues to depend on flat, prompt-retrieved tool descriptions. As the ecosystem of biomedical software expands, the link between the volume of available tools and context window limits results in inefficient execution, unstable planning, and tool confusion.

To address these challenges, we present BioManus, a biomedical agent designed natively for the Model Context Protocol (MCP). BioManus utilizes graph-scaffolded planning to manage structured biological capabilities. The system begins with the BioinfoMCP Compiler, a component that transforms disparate bioinformatics software into standardized MCP servers, thereby generating a comprehensive executable MCP ecosystem. Subsequently, this ecosystem is organized as a typed, heterogeneous MCP graph encompassing tools, operations, data types, and workflow stages.

During inference, BioManus identifies and retrieves compact, task-specific subgraphs, which are then used to synthesize scaffolds for operation-level workflows. This architectural approach separates the complexity of planning from the sheer size of the tool inventory. It achieves a context compression ratio of $\Theta(N / (h \cdot \bar{m}))$ under high-recall retrieval conditions, where $N$ represents the total number of tools, $h$ denotes the workflow horizon, and $\bar{m}$ (significantly smaller than $N$) indicates the average number of candidate tools per operation.

Evaluations on the BioAgentBench and LAB-Bench datasets demonstrate that BioManus outperforms advanced biomedical agent baselines in terms of context efficiency, workflow validity, and execution accuracy. These findings indicate a necessary paradigm shift: achieving scalable biomedical reasoning depends on utilizing structured, executable capability graphs rather than relying on increasingly large prompt-level tool retrievals.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC