Autonomous Drug Design via Workflow Agents: Integrating Chemical Semantics and Tool-Oriented Automation

Abstract Large language model (LLM)-based agents have shown promise in automating scientific tasks, yet existing systems for drug discovery often lack strategic, cross-step planning, and multi-objective coordination. Here, we present drugflow-skills, an LLM-driven agent framework that abstracts key drug design operations, including virtual screening, de novo molecule generation, molecular optimization, and multi-target design, into composable and reusable skills. The agent operates under a ReAct paradigm, executing iterative “Thought → Action → Observation” loops while interfacing with the DrugFlow backend. Across five representative tasks targeting ALK, ROS1, and IGF-1R, the agent demonstrates autonomous decision-making, cross-step planning, and multi-objective trade-offs, achieving improved predicted binding scores while reducing predicted toxicity and off-target liabilities. It also provides interpretable rationales and adapts strategies based on human feedback, enabling a human-agent co-optimization loop. These results support dynamic reasoning-driven drug design.