top of page
El Agente: LLM-Powered Multi-Agent System for Chemistry
- Overview: El Agente is an LLM-based multi-agent system designed to democratize computational chemistry by enabling users to perform complex quantum chemistry tasks via natural language interaction. It aims to lower technical barriers for both experts and non-experts in evaluating molecular properties and behavior.
- Architecture and Key Components: The system features a hierarchical agentic architecture for intelligent task distribution, automated error recovery, and performance improvement. Key components include task decomposition, adaptive tool selection, post-analysis, and autonomous file handling. It uses a global memory for shared context, agent-specific conversation history, and a grounding mechanism for environmental perception.
- Functionality: El Agente supports geometry optimizations, property predictions, and interfaces with various computational backends, high-performance computing job scheduling tools, and chemistry software. It provides transparent action trace exports for reproducibility and human oversight.
- Performance: Benchmarked on university-level course exercises and case studies, achieving an average task success rate of >87%. It demonstrates adaptive troubleshooting and error recovery during workflow execution.
- Implementation Details: Implemented in Python (v3.11.11), El Agente uses shell commands, the SLURM scheduler (v23.11.10), OpenBabel (v3.1.0), RDKit (v2024.09.5), Architector (v0.0.10), ORCA (v6.0.1), and xTB (v6.7.1).
- Roadmap: Future development includes incorporation of advanced computational simulations (DLPNO-CCSD(T), MC-PDFT, ADC(2), and NEVPT2), integration with PySCF and other platforms, adaptation for solid-state chemistry and materials science, support for periodic boundary conditions, uncertainty quantification, molecular dynamics (MD) simulations, integration with experimental databases, and integration with self-driving labs (SDLs).
Source:
bottom of page