Can Coding Agents Tackle Early-Stage Drug Discovery?

An evaluation tested the ability of three AI coding agents—Claude Code, Codex, and Gemini CLI—to perform early-stage drug discovery tasks in a specialized biomedical environment. The agents were benchmarked on 66 expert-curated tasks spanning structural reasoning, database screening, and patent mining. While no single agent dominated, key findings revealed that long-horizon planning remains the biggest challenge, with agents often failing to maintain context across multi-step problems. However, the agents executed well when provided with an expert's plan, indicating their primary weakness is in autonomous planning rather than task execution.