An auditable harness for biomolecular dynamics simulations by AI agents

AI agents now generate biological hypotheses faster than they can be tested. Hypotheses stated at the atomic level can be examined with molecular dynamics (MD) simulations. For new questions without a reference, trust in an agent-run simulation rests on a tool-written process record, not the agent's account. Existing MD agents leave this record to the language model. Here we present MDClaw, a harness whose validating tools and directed acyclic graph (DAG) check stages before execution and record every simulation stage. On MDDataBench, which grades outputs against deposited simulations, four open-weight models with MDClaw passed all three runs on 81-100% of 31 test systems. The DAG complements this outcome check. An ablation without the DAG revealed 11 successful runs that had skipped minimization or equilibration, or simulated a different structure. MDClaw carried questions from human and AI scientists through simulation to trajectories traceable to their inputs and settings.