Beyond Prompt-Based Planning: MCP-Native Graph Planning-based Biomedical Agent System

Researchers introduce BioManus, an AI agent system that uses graph-based planning and standardized Model Context Protocol (MCP) servers to automate biomedical workflows. The system addresses scalability challenges by organizing bioinformatics tools into structured capability graphs rather than relying on flat prompt-based retrieval, achieving significant improvements in execution accuracy and context efficiency.

BioManus represents a meaningful advancement in how AI agents handle complex, heterogeneous tool ecosystems. The core innovation lies in recognizing that as biomedical software environments expand, traditional prompt-based planning becomes inefficient—agents struggle with tool confusion and unstable execution when managing hundreds of incompatible bioinformatics tools. Rather than scaling by simply adding more tools to larger prompts, BioManus introduces a structural solution: the BioinfoMCP Compiler standardizes diverse bioinformatics software into MCP servers, then organizes them as a typed graph with clear relationships between tools, operations, data types, and workflow stages.

This approach fundamentally shifts how agents plan and reason. At inference time, the system retrieves only task-relevant subgraphs—dramatically compressing context size while maintaining high recall. The mathematical framing shows context compression scales as O(N / (h * m_bar)), meaning efficiency gains compound as tool inventories grow, unlike traditional methods that degrade with scale. Testing on BioAgentBench and LAB-Bench demonstrates measurable improvements in workflow validity and execution accuracy.

The broader significance extends beyond biomedical applications. This work suggests that scalable AI agent design requires structured, executable capability representations rather than semantic retrieval alone. For the AI infrastructure space, it highlights growing demand for standardization protocols like MCP that enable modular tool integration. The approach also has implications for other domains—finance, scientific research, and software engineering—where agents must coordinate complex, heterogeneous systems. As AI agents move from research to production workflows, this shift toward graph-scaffolded planning likely foreshadows how enterprise AI systems will be architected.

• →BioManus uses graph-based planning instead of flat prompt retrieval, enabling efficient scaling across hundreds of heterogeneous bioinformatics tools.
• →The BioinfoMCP Compiler standardizes incompatible bioinformatics software into compatible MCP servers, creating an executable tool ecosystem.
• →Task-specific subgraph retrieval achieves context compression with O(N / (h * m_bar)) efficiency, improving performance as tool inventories grow.
• →Experimental results show improvements in execution accuracy, workflow validity, and context efficiency compared to existing biomedical AI agents.
• →The work suggests a paradigm shift: scalable AI reasoning requires structured executable capability graphs, not just larger language model contexts.