AI agents now leave a tool-written audit trail for molecular dynamics: MDClaw passes all three runs on 81–100% of 31 test systems
Synopsis
AI agents can now propose atom-level biological hypotheses faster than they can be tested, but trust in an agent-run molecular dynamics simulation rests on a process record, and existing agents leave that record to the language model itself. MDClaw pairs validating tools with a directed acyclic graph (DAG, a graph recording which calculation stages depend on which) that checks stages before execution and records each one; on MDDataBench, which grades against deposited simulations, four open-weight models with MDClaw passed all three runs on 81–100% of 31 test systems. An ablation without the DAG found 11 runs that passed every output check yet had skipped minimization or equilibration, or simulated a different structure.
Interpretation
A molecular dynamics result produced by an agent can now be traced independently of the agent's own account: each calculation stage is written into a directed acyclic graph by the tools, and a trajectory can be followed back through parent nodes to its inputs and settings. Earlier molecular dynamics agents left the order of stages and the passing of files to the language model, so the record of what ran was mainly the agent's own log and summary, which depends on the same agent it is meant to check. Across 30 non-membrane test systems, four models and three runs per system with tools only, 11 of 334 successful runs without the DAG had process errors, versus none of 343 successful runs with the DAG.
The approach writes expert protocols as 14 skill documents and 100 command-line tools that validate inputs before computing and register each calculation as a node in the graph. A node runs only when its parent stage is complete, the stage transition is allowed, inputs come from recorded ancestors, and execution settings agree with declared conditions, so a file produced outside the record cannot enter a later stage. On the 31 MDDataBench test systems, tools alone gave pass3 on 25–29 systems (81–94%) across the four models, and tools plus skills gave 28–31 systems (90–100%); without MDClaw, only 21 of 372 individual runs passed.
A rejected calculation does not force a restart: the agent can branch from the last valid node and continue after a correction. Tools return a structured error code with a specific remedy, and failed and superseded nodes stay in the record. In ligand task 5OD1, a tool rejected an invalid chemical string, a second preparation reported charge zero against the prompt's required −2, and a third from the same structure passed all 20 checks; in membrane task 6ME3, an equilibration node declared at 310 K but configured at 300 K was rejected before simulation, and the agent branched a 310 K node from the completed minimization node that passed all 20 checks.
Perspective
The demonstrated scope is system preparation and 1–2.5 ns short trajectories on the 31 MDDataBench test systems (eight system classes, including complexes, nucleic acids and one membrane protein), and grading measures conformity to a defined simulation task rather than equilibrium convergence or biological validity; the application examples cover Abl1 substrate recognition, Abl1–cyclotide complexes and five PNG-1 hypotheses proposed by Biomni.
The DAG makes every step inspectable but is not tamper-proof, because an agent with write access could alter its records, and it covers only calculations run through the tools; the DAG also cannot judge whether a force field, starting model or collective variable suits a scientific question. Reliable execution does not overcome the sampling limits of molecular dynamics: in the PNG-1 study only one of five hypotheses was partially supported within the time limit, and the free-energy errors come from three replicas per leg and do not establish convergence.