- The paper surveys 49 computational-chemistry AI agents and finds rapid growth from 4 systems in 2024 to 33 in the first seven months of 2026, with most progressing from single calculations toward multi-step in silico experiments.
- The paper identifies broad coverage of structure generation, molecular dynamics, and electronic-structure calculations but limited treatment of reaction mechanisms and ensemble free energies, while reported Aitomia reliability falls from 99.6% for low-autonomy tasks to 45.6% at higher autonomy.
- The paper argues that specialized chemistry agents may be intermediate technologies as general-purpose coding agents absorb their skills and tools, while licensing gaps, weak testing, limited deployability, and the absence of demonstrated campaign-level autonomy remain major barriers.
Scope and survey methodology
This Perspective by Dral, Nawaz, and Ullah presents a systematic survey of agentic systems for computational chemistry, defined as systems whose object is an atomistic simulation — quantum-chemical, molecular dynamics, or machine-learning-potential calculations on molecules, materials, or catalytic surfaces — set up, run, and interpreted by AI. The survey covers 49 systems with a literature cutoff of 8 August 2026, ordered by first public appearance. The corpus was assembled from reference lists of prior surveys, arXiv and ChemRxiv listings, and colleague suggestions, so the authors explicitly note it is not the output of a single reproducible query and cannot estimate their own recall; the survey grew from 37 to 49 systems during preparation, making 49 a lower bound. Two evidentiary standards are applied separately: a scientific task counts where authors report their system performing it on their own account, while delegation is held to the stricter standard of what a paper demonstrates in its body, excluding future-work aspirations.
Growth and task coverage
The growth is rapid: 4 systems appeared in 2024, 12 in 2025, and 33 in the first seven-plus months of 2026. The authors characterize this as an explosion, and the trajectory is consistent with broader agentic-AI trends. Task coverage is uneven. Structure generation — "task zero" of any workflow — is nearly universal, ranging from 3D embedding of SMILES strings to generation and screening of hundreds of hypothetical crystals. Molecular dynamics and electronic-structure calculations are widely represented, as is periodic DFT. By contrast, coverage of reaction mechanism is thin: transition states and reaction barriers, described as the core business of mechanistic computational chemistry, are handled by only a handful of systems, which the authors attribute to the intrinsic difficulty of the task.
Free energies appear in 12 systems, but the split is diagnostic. Eight use the static route — harmonic frequencies and ideal-gas entropies converted to Gibbs energies — yielding pKa​ values, ring strains, and adsorption free-energy diagrams at the cost of a frequency calculation. Only four sample the ensemble, via umbrella-sampling potentials of mean force, endpoint MM/PB(GB)SA, or alchemical perturbation with Bennett and multistate-Bennett estimators. The survey therefore covers structures and dynamics broadly but reaction thermodynamics and kinetics only shallowly.
Delegation scale and autonomy
The authors classify autonomy on a five-position scale: single calculation, in silico experiment (planning, executing, and analyzing a sequence of calculations), paper level, campaign level (multiple papers within a scoped project), and agenda level (self-defined research direction with no human intervention). The trend over two and a half years is clear: early systems operated at the scale of a single calculation, the bulk has shifted to the experiment level, and only a few systems claim paper-level autonomy — CatMaster, Protomia, and MDAgent — without clear evidence of producing papers en masse without human supervision. No surveyed system demonstrates campaign-level autonomy, and the authors judge agenda-level discussion premature. They note, however, that public reports lag frontier-lab activity and that campaign-level efforts are actively pursued in their own labs. The margin between the two commonest delegation positions is fragile: under one plausible recoding, 23 single-calculation systems stand against 24 in-silico-experiment systems, so the ordering survives by one system rather than three.
Architectural evolution
Three architectural generations are distinguishable. Early systems shipped predefined tool functions written before the run, with the LLM reduced to choosing tools and parameters. Later systems increasingly rely on agent-written code composed and executed during the run, the Model Context Protocol (MCP) for calling unshipped tools, and retrieved skills libraries — written procedures the agent follows in place of function calls. The orchestration layer moved in parallel: LangGraph was the backbone of early architectures, now displaced by loose collections of skills atop general-purpose agents such as Claude Code. Importantly, the newer layers accumulate beside the predefined one rather than replace it, since skills often contain predefined functions. This architectural shift enabled closed-loop operation — letting a result change what runs next — which is now the norm rather than the exception.
A structural dependency runs through this evolution: state-of-the-art underlying models are built by private companies whose goals are not necessarily aligned with autonomous scientific research, so the field's progress is bounded by repurposing rapidly changing external tools, causing a significant development lag.
Reproducibility and availability
Availability is a major weakness of the field. Only 27 of the 49 systems state a licence (most often MIT; the authors' own Aitomia is Apache-2.0). Of the 22 with no licence found, 6 are open repositories carrying no licence file, leaving readers code they can read but may not legally reuse. Most systems ship no tests and no continuous integration, and several cannot run as published due to missing modules, commented-out entry points, or hard-coded author-machine paths. Only 4 systems can be tried online without installation: TritonDFT, VASPilot, AutoSolvateWeb, and the authors' Protomia (a hosted service, not open source). Evaluation is equally problematic: benchmarking agentic systems requires scarce human judgement, the pace of architectural and model change makes equal-footing comparison practically impossible, and any benchmark, once leaked, contaminates subsequent evaluations.
On robustness at the single-calculation level, the authors report concrete numbers from their Aitomia benchmark: a 99.6% success rate for low-autonomy single computational tasks, dropping to 70.9% and 45.6% as demanded autonomy increases — a steep autonomy–reliability trade-off. Those figures reflect the obsolete LangGraph-based architecture; the successor Protomia offers experiment- and paper-level autonomy (demonstrated on a Diels–Alder reaction energy calculation and LaTeX mini-paper generation with a deterministic provenance auditor) but still requires substantial human supervision at the ideation-to-paper scale, and awaits formal evaluation and publication.
Commoditization by generalist agents
A central and deliberately provocative claim is that general-purpose coding agents (Claude Code, GitHub Copilot, Codex, Qoder, and emerging open alternatives) are commoditizing the construction of specialized systems. The same tools that collapsed the barrier to building agentic systems raise the barrier to their adoption: once a specialized system reduces to skills and MCP tools, a general-purpose agent can call them directly, and general-purpose agents improve faster than any specialized system in the survey. The authors predict that once general-purpose agents match specialized capabilities and support autonomous research (e.g., Claude Science), specialized systems will have little reason to exist. This is supported by their own adoption data: uptake of their freely available systems was slower than expected, and members of their own group — despite repeated demonstrations of Protomia's capabilities — still copy-paste scripts to online GPT chats and run them manually on HPC clusters, or prefer general-purpose coding agents. The authors conclude that if the people closest to a free, capable system reach for something else, the binding constraint on the field is not what the systems can do.
The paper is itself an existence proof of the trend: it was substantively produced by the "AI Dral Group," an AI research-agent collective operating on Claude Code with the Claude Opus 5 model, with Protomia used for literature retrieval and citation verification. Figures are generated by scripts fixed to the database, and the authors take full responsibility for verified content. The authors also flag a conflict of interest: two surveyed systems are their own, one of which (Protomia) was used in preparing the article, and P.O.D. holds equity in Aitomistic.
Limitations and open questions
The Methods section is unusually candid about three limits that could move the reported numbers. First, the authors read papers rather than ran systems, so every count measures what a paper shows; a system that works better than it writes is undercounted. Second, recall cannot be estimated, so 49 is a floor — indeed, the count rose during preparation. Third, the delegation placements are the authors' own with no independent recoding, and the single-calculation versus in-silico-experiment ordering is fragile, as noted above. The task-counting rule (implemented capability counts even without a worked example) is generous relative to the stricter delegation standard, which the authors acknowledge by stating the two standards separately.
Open questions the paper leaves explicit: whether paper-level autonomy can be demonstrated at scale without human supervision; whether campaign-level systems exist behind closed doors at frontier labs; whether specialized systems can survive generalist commoditization; and how agentic systems can be evaluated fairly given benchmark contamination, evaluation cost in tokens, and the pace of architectural change. The authors also concede they have "no answer" to what established specialists, teachers, and students should spend their efforts on given the speed of disruption — a stated limitation of the Perspective's prescriptive value.
Conclusion
The survey documents a field that grew from 4 to 49 agentic systems in two and a half years, with delegation shifting from single calculations to full in silico experiments, architectures migrating from predefined tool functions toward agent-written code, MCP, and skills atop generalist runtimes, and paper-level autonomy claimed but not yet demonstrated at scale. Coverage of structures and dynamics is broad; reaction mechanisms and ensemble free energies remain thin. Reproducibility — licensing, testing, deployability — lags badly behind capability claims. The paper's strongest claim, supported by the authors' own adoption experience, is that specialized agentic systems for computational chemistry are an intermediate form likely to be absorbed by general-purpose coding agents. Whether the endpoint is a fully autonomous computational chemist, and on what timescale, remains an open empirical question that no system in this survey yet answers.