BioMaster: Multi-agent system for automated bioinformatics analysis workflow

The growing volume and complexity of biological data have made bioinformatics workflows increasingly labor-intensive, error-prone, and difficult to scale. Large language model-based agents offer potential for automation but often fail in complex, multi-step analyses because of limited robustness. We present BioMaster, a multi-agent framework that integrates workflow planning, execution, error recovery, and output validation. BioMaster incorporates a dual retrieval-augmented design to leverage domain knowledge for tool selection, parameterization, and adaptation across tasks. A dedicated debug agent supports real-time error detection and correction, while memory optimization enables long, multi-stage workflows. In benchmarking across 49 bioinformatics tasks spanning 102 tools, BioMaster completed substantially more workflows than did existing automated systems, particularly in complex, interdependent pipelines. BioMaster supports both proprietary and open-source language models, enabling flexible deployment across different computational settings.