BioManus: Reshaping Biomedical Workflow with Graph-Scaffolded Planning

Biomedical agents are poised to revolutionize biological workflows, yet they've hit a bottleneck. Current systems buckle under the weight of heterogeneous bioinformatics tools and flat prompt-retrieved tool descriptions. Enter BioManus, an MCP-native agent offering a fresh solution.

Graph-Scaffolded Planning

BioManus is built on graph-scaffolded planning over structured biological capabilities. It introduces the BioinfoMCP Compiler, which standardizes diverse bioinformatics software into MCP servers. This creates a comprehensive and executable MCP ecosystem, transforming how biomedical tasks are approached.

But why does this matter? The tool ecosystem is organized into a typed heterogeneous MCP graph. This graph spans across tools, operations, datatypes, and workflow stages. At inference time, BioManus isn't overwhelmed by the sheer number of tasks. Instead, it retrieves compact, task-specific subgraphs and synthesizes operation-level workflow scaffolds. The result is a decoupling of planning complexity from raw tool inventory size.

Efficiency Gains

The paper's key contribution: BioManus achieves a context compression ratio, Theta(N / (h * m_bar)), under high-recall retrieval. Here, N represents the total tool count, h is the workflow horizon, and m_bar is the average number of candidate tools per operation. This isn't just a technical marvel. It signifies a shift in how we think about scalable biomedical reasoning.

Experiments on BioAgentBench and LAB-Bench show that BioManus doesn't just hold its own against advanced biomedical agent baselines. It improves execution accuracy, workflow validity, and context efficiency. That's a bold statement in a field that's growing more complex by the day. Why keep relying on outdated methods when a structured executable capability graph offers clarity and scalability?

Looking Ahead

This builds on prior work from the field of biomedical agents, but BioManus takes a leap forward. It suggests that we shouldn't just expand prompt-level tool retrieval. Instead, structured capability graphs are the way forward for scalable reasoning.

In an age where complexity can stifle innovation, BioManus offers a way to cut through the noise. Will the biomedical field embrace this paradigm shift? The evidence suggests it should. BioManus isn't just another tool. It's an evolution in how we approach bioinformatics workflows.

Code and data are available at the project's repository, offering a chance to dive deeper into this transformative approach.