MolClaw: An Autonomous Agent with Hierarchical Skills for Drug Molecule Evaluation, Screening, and Optimization

A team of researchers from multiple institutions (including Lisheng Zhang, Lilong Wang, Xiangyu Sun, and 13 others) has introduced MolClaw, an autonomous AI agent designed to tackle the complex workflows of computational drug discovery. The agent addresses a critical gap: current AI agents struggle to maintain robust performance when orchestrating dozens of specialized tools in multi-step workflows. MolClaw unifies over 30 specialized domain resources through a novel three-tier hierarchical skill architecture comprising 70 total skills. At the tool level, skills standardize atomic operations; at the workflow level, they compose validated pipelines with quality checks and reflection; and at the discipline level, a scientific principle governs planning and verification across all scenarios.

The researchers also introduced MolBench, a comprehensive benchmark featuring molecular screening, optimization, and end-to-end discovery challenges that require 8 to over 50 sequential tool calls. MolClaw achieves state-of-the-art performance across all metrics in this benchmark. Critically, ablation studies reveal that performance gains are concentrated on tasks demanding structured workflows, while vanishing on those solvable with ad-hoc scripting. This establishes workflow orchestration competence as the primary capability bottleneck for AI-driven drug discovery. The paper spans 59 pages with 10 figures, and code and data are promised for release.