Can AI Agents Do the Work of Drug Discovery?

Scale Labs and Phylo developed DrugDiscoveryBench to measure how well AI agents perform early-stage drug discovery tasks like looking up molecular structures and mining patents. The benchmark found that top agents solve about half the problems, excelling at short tasks but failing on long, multi-step workflows where they lose the original objective. Agent performance significantly improves when provided with an expert-created, step-by-step plan or when the underlying agent system, or harness, is optimized. This suggests that combining AI's ability to handle manual data work with human expertise and better tooling is the key to accelerating scientific discovery.