El Agente Q

Purpose and contribution

El Agente Q applies the El Agente cognitive architecture to computational and quantum chemistry. The study describes a multi-agent system that translates natural-language tasks into computational workflows, coordinates calculations, and interprets their results.

Its contribution is an evaluated orchestration approach for computational chemistry, rather than a new electronic-structure method or a demonstrated fully autonomous wet laboratory. Read the platform and research context →

Scientific team and attribution

Prof. Varinia Bernales and Prof. Alán Aspuru-Guzik are coauthors and corresponding authors of the journal article; Bernales is its designated lead contact. The university announcement describes Bernales as Matter Lab’s director of research and Aspuru-Guzik as its principal investigator, as of May 2, 2025.

The work is credited to the full multidisciplinary author team below. Corresponding-author and lead-contact roles do not establish sole conception, company founding, or responsibility for every implementation contribution.

Publisher’s author and contact information · Dated institutional announcement

Peer-reviewed publication

Yunheng Zou; Austin H. Cheng; Abdulrahman Aldossary; Jiaru Bai; Shi Xuan Leong; Jorge Arturo Campos-Gonzalez-Angulo; Changhyeok Choi; Cher Tian Ser; Gary Tom; Andrew Wang; Zijian Zhang; Ilya Yakavets; Han Hao; Chris Crebolder; Varinia Bernales; Alán Aspuru-Guzik. (2025). El Agente: An autonomous agent for quantum chemistry. Matter, 8(7), 102263. Published July 2, 2025.

DOI: 10.1016/j.matt.2025.102263

Related preprint: arXiv:2505.02484, version 2 (August 8, 2025). The preprint and journal article are related versions of the work, not separate research systems.

Evaluation and limitations

The abstract reports an average task success rate above 87% across six university-level exercises and two case studies. This is a task-specific benchmark result, not a universal reliability or chemical-accuracy guarantee.

In the version 2 preprint, exercises were evaluated at two difficulty levels with ten repetitions per exercise and identical prompts. Prompts specified the density functional and basis set; success was assessed using task-specific rubrics. Results depend on the configured models, tools, prompts, and orchestration.

Reported limitations include missing final reports, unverified assumptions or approximations, unstable generated geometries, and coordination failures requiring tuning or follow-up prompting. Calculation logs and expert review remain important.

Full methods, benchmark tables, and limitations in the preprint. These results do not support an unqualified “10X–100X” speedup claim.

Research data and software availability

The preprint’s availability statement links to research data for evaluating the reported conclusions. This is a research-data link, not a verified source-code release.

A public source-code repository, license, and tagged release have not been verified for this page. For project access, see El Agente; for institutional resources, see Matter Lab’s Open Software & Data directory.