An auditable harness for biomolecular dynamics simulations by AI agents

MDClaw, an auditable harness for molecular dynamics, enabled four open-weight AI models to pass all three MDDataBench runs on 81–100% of 31 test systems.

• Why it matters: Agent-run MD can validate atomic-level hypotheses, but trust is limited when the process record is left to the language model rather than tool-verified execution.
• What they did: MDClaw uses validating tools plus a directed acyclic graph (DAG) that checks stages before running and records every simulation stage on MDDataBench.
• The result: Without the DAG, an ablation produced 11 successful runs that skipped minimization or equilibration or used a different structure, while MDClaw produced trajectories traceable to inputs and settings.

AI summary of the abstract · may contain errors · not medical advice · read the paper.