Paper Overview
- Field: Machine Learning
- Authors: Zhangtianyi Chen, Florensia Widjaja, Wufei Dai
- Published: 2025-06-01
- arXiv: 2606.04494
Summary
Biomedical agents promise to automate complex biological workflows, yet current systems face two fundamental bottlenecks: bioinformatics tools are highly heterogeneous in interfaces and execution environments, while agent planning still relies on flat prompt-retrieved tool descriptions. As biomedical software ecosystems grow, the coupling between tool coverage and context size leads to tool confusion, unstable planning, and inefficient execution.
The authors introduce BioManus, an MCP-native biomedical agent built on graph-scaffolded planning over structured biological capabilities:
1. BioinfoMCP Compiler — converts heterogeneous bioinformatics software into standardized MCP servers, yielding a large executable MCP ecosystem. 2. Typed heterogeneous MCP graph — organizes the ecosystem over tools, operations, data types, and workflow stages. 3. Subgraph retrieval at inference — retrieves compact task-specific subgraphs and synthesizes operation-level workflow scaffolds.
This design decouples planning complexity from raw tool inventory size, achieving a context compression rate of Theta(N / (h * m_bar)) under high-recall retrieval, where N is the total number of tools, h is the workflow scope, and m_bar (much smaller than N) is the average number of candidate tools per operation.
Experiments on BioAgentBench and LAB-Bench show BioManus outperforms state-of-the-art biomedical agent baselines in execution accuracy, workflow validity, and context efficiency. The work suggests a paradigm shift: scalable biomedical reasoning requires structured executable capability graphs rather than ever-larger prompt-level tool retrieval.
*Auto-collected on 2026-06-05.*