DrugHarness is an evolvable agent harness designed to address the need for composable scientific operations and adaptive constraint handling in computational drug discovery. Its environment bundles more than 80 scientific functions, biomedical resources, and task‑specific code, preserving links among candidate molecules, experimental outcomes, and generated artifacts.
After each execution trajectory, the system derives action hypotheses from the feedback, specifying observable applicability conditions, recommended actions, and supporting evidence. An independent review agent then validates these hypotheses through source verification and, when necessary, targeted execution trials. Only skills that pass this review are admitted to subsequent tasks and are continuously refined by new feedback.
Evaluation on the MolBench and SMDD‑Bench benchmarks shows strong performance. Without evolution, DrugHarness surpasses reported MolClaw‑CC results by 10.8 percentage points in affinity‑comparison accuracy and by 55 % in mean top‑three active hits. With evolution, Qwen’s affinity‑comparison accuracy rises from 62.2 % to 81.1 % and optimization success improves from 79.5 % to 87.2 %. In a prospective HPK1 study, the system yields candidates with a more favorable median docking score than Biomni, and a synthesized compound achieves a cellular IC₅₀ of 14.19 nM—a 2.64‑fold potency gain over the best reference and comparable activity to two compounds designed by seasoned medicinal chemists.
These findings support conditional skill evolution as a powerful mechanism for scientific agents and demonstrate a workflow that bridges computational drug design with experimental validation.
Review: DrugHarness illustrates how dynamic skill evolution can drive efficient screening and optimization, offering a reproducible blueprint for future AI‑powered drug discovery efforts.