Transparent reporting for agentic catalysis enabled by artificial intelligence: Community guidelines and a publication checklist
Hongliang Xin, John R. Kitchin, Núria López, Neil M. Schweitzer, Milad Abolhasani, Nongnuch Artrith, Líney Árnadóttir, Kamal Choudhary, Rui Ding, Anatoly I. Frenkel, Joseph A. Gauthier, Bryan R. Goldsmith, Amir Barati Farimani, Lars C. Grabow, G.T. Kasun Kalhara Gunasooriya, Guoxiang Hu, Tyler R. Josephson, Heather J. Kulik, Ritesh Kumar, Teodoro Laino, Hao Li, Xiao-Yan Li, Wanlu Li, Suljo Linic, Chong Liu, Cong Liu, Fudong Liu, Mingjie Liu, Piao Ma, Andrew J. Medford, Sukrit Mukhopadhyay, Pengfei Ou, Christopher Paolucci, Jiayu Peng, Cory Phillips, Marc D. Porosoff, Long Qi, Andrew S. Rosen, Shijing Sun, Tibor Szilvási, Johannes Voss, Siwen Wang, Xiaonan Wang, Kirsten Winther, Qin Wu, Di Zhang and Zisheng Zhang, “Transparent reporting for agentic catalysis enabled by artificial intelligence: Community guidelines and a publication checklist”, Chem Catalysis 6(8), 101755 (2026). https://doi.org/10.1016/j.checat.2026.101755
This summary was written by a language model from the paper and has not been checked line by line. Rely on the published paper for details.
This is a Perspective that proposes reporting guidelines; it reports no new experiments or calculations. The summary is based on the author proof of the article.
Question
When AI systems in catalysis do more than make predictions, and instead plan, choose and run experiments or simulations in a closed loop, what has to be reported for others to judge and reuse the work (Xin et al. 2026)? Heterogeneous catalysis already struggles with rigor and reproducibility because outcomes depend on catalyst preparation, the evolution of active sites under reaction conditions, and reactor transport. Agentic workflows add further hidden variables: prompts, decision policies, tool access, safety constraints and the extent of human intervention. The authors also argue that reproducibility must be redefined for such systems, since an agent’s stochastic step-by-step trajectory cannot and need not be repeated exactly.
Approach
A large group of catalysis and AI researchers, conceived by Xin, Kitchin, López and Schweitzer and growing out of an NSF-funded workshop on AI in heterogeneous catalysis, proposes TRACE-AI (transparent reporting for agentic catalysis enabled by artificial intelligence). It is positioned against existing frameworks: the FAIR data principles, best practices for ML in chemistry, the clinical SPIRIT-AI and CONSORT-AI extensions, and the cross-disciplinary REFORMS recommendations. TRACE-AI is organized around the research life cycle in three coupled domains (data management, model development and workflow integration) linked to the campaign’s scientific objectives, and is accompanied by a glossary and a publication checklist. The checklist and templates for agent cards, data manifests and autonomy run logs are maintained as a versioned, CC BY 4.0 resource on GitHub; the article references release v2.1.0.
Key findings
The central reframing is that reproducibility for agentic campaigns means end-to-end traceability plus statistical consistency: independent runs under comparable constraints should reach similar regions of design space, performance distributions or mechanistic conclusions, and every realized trajectory should be logged even though it will not be repeated exactly. The glossary separates trajectory traceability from exact trajectory reproducibility, and counts scripted automation as an agent only when action selection adapts to observations.
The paper defines autonomy levels by analogy to autonomous driving: A0 (manual), A1 (decision support, human selects among AI proposals), A2 (supervised autonomy, human approves each execution), A3 (conditional autonomy within pre-approved bounds), A4 (continuous autonomy with only interlocks and audit logs), and A5 (distributed autonomy across agents or laboratories). Authors are asked to declare the level, including changes within a study.
It introduces the agent card, a versioned description of the observation space, action space, objectives, constraints, policy (for example Bayesian optimization, active learning, reinforcement learning or heuristics) with hyperparameters, tool access, safety checks and sources of stochasticity. For LLM components, authors should give model identity and version, inference settings such as temperature and context window, accessible data and tools, and the prompt strategy (verbatim when it determines the outcome).
The checklist has 18 items in four groups. Data management (A1 to A5) covers harmonizing multi-source data, code that regenerates the ML-ready dataset, descriptor definitions, dataset identifiers and metadata, and a policy on negative or failed experiments. Model development (B1 to B5) covers full model specification, leakage-preventing splits, metrics with uncertainty rather than best-only results, justified metrics and multi-objective trade-offs, and calibrated uncertainty quantification. Workflow integration (C1 to C5) covers agent policy, autonomy and human interventions, stopping and failure handling, safety bounds and interlocks, and a traceable trajectory record. Executable artifacts (D1 to D3) covers tutorials or notebooks linked to the manuscript and documented execution environments.
Specific recommendations include grouping near-duplicate compositions and shared precursors or batches in one split, prospective validation for discovery claims, distinguishing epistemic from aleatoric uncertainty, recording manual changes to exploration settings as interventions, specifying stopping rules in advance, and reporting how often available tools are actually invoked. Quality-control steps such as replicates and reference catalysts can be built into the agent’s action space and logged.
Significance
TRACE-AI translates the group’s long-standing advocacy for data sharing and reproducible ML reporting into a concrete community standard for autonomous catalysis, aimed at authors, reviewers, editors and funders. It anticipates networks of cloud-connected laboratories where shared standards, benchmarks and decision logs make agentic studies comparable.