Artificial intelligence (AI) agents and large-language models (LLMs), such as the model underpinning OpenAI’s conversational platform ChatGPT, are now widely used by people worldwide, both in informal and professional settings. Over the past decade or so, some of these models have also been adapted to tackle complex research problems rooted in various fields, including biology, physics, medical sciences and chemistry.

Existing computational tools employed by chemists are often highly sophisticated and complex. Their complexity makes them inaccessible to non-expert users and often even difficult for expert chemists to use.

Researchers at Matter Lab at the University of Toronto and NVIDIA have developed El Agente Q, a new LLM-based system that could allow chemists, particularly those specialized in quantum chemistry, to easily generate and execute quantum chemistry workflows, sequences of computational tasks required to study specific chemical systems at the quantum mechanical level.

The new agent, introduced in a paper published on the arXiv preprint server, relies on a new multi-agent cognitive architecture that can effectively decompose a task, select computational tools that can help solve it and perform relevant analyses.

“El Agente Q resulted from our team’s realization that even though quantum-chemistry simulations are crucial for accelerating molecular and materials design, executing them is still difficult and inaccessible to anyone not an expert in the field,” Alán Aspuru-Guzik, Senior Director of Quantum Chemistry at NVIDIA and director of the University of Toronto’s Acceleration Consortium and the Matter Lab, told Phys.org.

“Using large language models, we created an agent system that partners with scientists in quantum chemistry to accelerate scientific progress.”

El Agente Q, the new LLM-based system for tackling quantum chemistry problems, is comprised of over 20 AI agents. Each of these agents is “specialized” in specific quantum-chemistry tasks, possesses its own working memory and executes its own actions.

“These sub-agents are arranged for efficient collaboration, and can decompose tasks, report progress, and resolve exceptions,” explained Yunheng Zou, lead author of El Agente. “This design enables strong instruction following, reduced context, and guided reasoning on domain-specific tasks. In six university-level exercises, El Agente Q solved 87% of its test problems in only one try.”

To demonstrate the potential of El Agente Q, the researchers tested it on six university-level course exercises and two quantum chemistry-related case studies. Remarkably, they found that the system could successfully solve 87% of the tasks used in their experiments, while also identifying and fixing errors during the execution of tasks in real-time.

“We demonstrated that El Agente Q can automate a large portion of quantum chemistry tasks in a user-friendly way, making it easier to understand materials and molecular properties and behavior,” said Varinia Bernales, research director of the Matter Lab. “This lets more scientists access both molecular and materials science fields through AI, paving the way to greater discoveries.”

The new LLM-based system developed by the Matter Lab led by Aspuru-Guzik is both better performing and easier to use than many existing computational tools for quantum chemistry research. The researchers will soon release El Agente Q to the public and hope that it will prove to be a useful tool for both chemistry researchers and students.

“In the coming months, we will launch a cloud-enabled alpha version of El Agente Q on www.elagente.ca, which will give users direct access to its LLM-driven quantum-chemistry workflows via plain-English prompts,” added Aspuru-Guzik.

Bernales, on the other side, says “Beyond that, we’re looking to apply it to a broader set of computational tasks, improve user interaction, and prepare for easy integration with laboratory systems. We’re also looking to continue creating more capable scientific agents.”

© 2025 Science X Network