Drug discovery spans molecular design, property prediction, and virtual screening, forming a tightly coupled workflow with interdependent computational tasks and a plethora of specialized tools. Coordinating scientific models and computational resources under such complexity remains a major hurdle for existing large language model (LLM) based agents, which struggle with autonomous planning, adaptive model selection, and reliable multi‑stage execution.\ \ To address these issues, we introduce DrugCoding, a scientific programming agent that unifies task planning, adaptive model routing, multi‑agent collaboration, and programmatic tool execution. DrugCoding autonomously constructs research plans, dynamically invokes domain‑specific models, generates and runs code according to task requirements, and leverages multiple agents to ensure efficient execution and robust validation.\ \ Key components are:\
- Task planner – translates natural‑language objectives into a multi‑step workflow.\
- Model router – selects the most suitable specialized model (e.g., molecular generator, property predictor) for each sub‑task.\
- Code generation & execution engine – produces tool‑compatible code snippets and runs them in a sandboxed environment.\
- Multi‑agent collaboration layer – distributes sub‑tasks among agents, aggregates results, and performs consistency checks.\ \ In evaluation, we conducted multi‑turn tool‑calling experiments across various language models and scientific tool interfaces. DrugCoding achieved an average session‑level Pass@1 of 97.35% on the full benchmark, surpassing the strongest baseline Codex by 9.24%. On a challenging subset, Pass@1 reached 97.92%, again outperforming the baseline.\ \ These findings demonstrate DrugCoding’s strong capability in scientific tool invocation and multi‑stage task execution, offering a reliable automation solution for high‑complexity domains such as drug discovery.\ \ Review: DrugCoding’s integration of planning, adaptive modeling, and collaborative agents delivers a highly automated pipeline for complex drug discovery workflows, markedly improving tool‑calling success rates and paving the way for more robust LLM‑driven scientific automation.