Science Done on a Machine by a Machine: AI Agents in Computational Chemistry
Pavlo O. Dral, Hassan Nawaz, Arif Ullah
physics.chem-ph, cs.AI, physics.comp-ph
2026-08-19
A survey of 49 computational-chemistry agents through 8 Aug 2026: four in 2024, 33 in 2026. Structure and MD are common, mechanisms rare; most are used mainly by their authors.
Single computational-chemistry jobs have been scriptable for years. Scripts did not remove the expert; they moved the work to choosing what to run, writing inputs, watching failures, revising the plan, and reading outputs. After GPT-4, building agents for atomistic simulation became a field. Dral’s group at Xiamen, Anhui University, and Aitomistic survey systems whose object is a quantum-chemical, molecular-dynamics, or machine-learned-potential calculation on molecules, materials, or catalytic surfaces, through 8 August 2026: 49 entries. The point is not another agent. It is that specialized systems may already be losing the plot.
This is a perspective, not a methods paper. Inclusion is an AI system that sets up, runs, and interprets an atomistic calculation. The authors plot first public appearance, then cross-tabulate task type, orchestrator, autonomy grain, licence, and whether a reader can try it without installing anything. Autonomy is scored from a single calculation, through an in silico experiment, through a paper, up to an unsupervised campaign and a self-set agenda. Their own Aitomia and later Protomia are marked on the figure; teaching and in-group use are participant observation, not a third-party blind test.
Public count: 4 systems in 2024 (0.3 per month), 12 in 2025 (1.0), 33 by 8 August 2026 (4.6). Task coverage is structure generation 33, molecular dynamics 27, electronic structure 15, periodic DFT 15, vibrations 14, free energy 12, screening 9, transition states 9, excited states 8, MLIP development 6. Eight of the free-energy systems take the static harmonic plus ideal-gas route; four sample the ensemble. Transition states and barriers remain thin, which matches how hard that job is.
Autonomy: 21 at calculation scale, 24 at experiment scale, 4 claiming paper scale, 0 at campaign or agenda. Every reported system still keeps a human in the loop. Early orchestrators cluster on LangGraph (9) and LangChain (6); 12 write their own loop. In 2026, systems start sitting on Claude Code, Codex, and OpenClaw. Newer stacks add agent-written code, MCP, and skills; predefined tools remain, they are no longer the only interface.
Reuse looks worse. 27 of 49 state a licence; 22 have none, including 6 public repos with no licence file. Most ship no tests or CI; some miss modules, comment out the entry point, or hard-code a path from the author’s machine. Only four can be tried online without installing: TritonDFT, VASPilot, AutoSolvateWeb, and the hosted Protomia. Aitomia hit 99.6% on low-autonomy single tasks, then 70.9% and 45.6% as autonomy rose, on an obsolete LangGraph stack. Protomia claims experiment- and paper-level ability; a formal eval is still unpublished.
For people building AI-for-science tools, the cold reading is that specialized chemistry agents are being commoditized by general coding agents plus skills and MCP. The explosion in count and the thin adoption outside author groups can be true at once: the bottleneck moved from “can it compute” to “will anyone change their workflow.” Single calculations are already robust, experiment-scale is becoming the bulk, paper-scale still wants heavy human review. If the job is a standard protocol, ask whether an existing coding agent can call your skills before minting another branded orchestrator.
The cutoff is 8 August 2026; public write-ups lag internal lab work. Task labels come from papers, not a shared blind benchmark, so systems cannot be ranked. Evaluation needs humans, burns tokens, and contaminates easily; the authors say a fair 49-way bake-off is practically impossible. Aitomia numbers and Protomia capabilities come from this lab; slow uptake, including group members still pasting scripts into a web chatbot, is anecdotal. The closing line that everyone is digging their own grave is a stance, not a falsifiable result. The manuscript itself was drafted with the lab’s AI research agents and finished by Dral.