AGAPI-Agents: An Open-Access Agentic AI Platform for Accelerated Materials Design on AtomGPT.org

Area of Science:

• Materials Science
• Artificial Intelligence
• Computational Chemistry

Background:

• Agentic AI systems are increasingly connected to external scientific tools.
• The impact of tool access on prediction accuracy for these systems is not well understood.
• Existing methods lack comprehensive evaluation of agentic AI in scientific applications.

Purpose of the Study:

• To present AGAPI (AtomGPT.org API), an open-access platform integrating large language models (LLMs) with scientific tools.
• To evaluate the impact of tool integration on the prediction accuracy of agentic AI systems.
• To demonstrate autonomous multi-step workflows for scientific discovery.

Main Methods:

• Integration of eight open-source LLMs with 18 REST endpoints (28 agent tools, 50 web apps).
• Utilized materials databases, force fields, tight-binding band structures, X-ray diffraction, and protein structure tools.
• Employed a three-evaluation residual decomposition on JARVIS-Leaderboard electronic-structure test sets.

Main Results:

• AGAPI reproduces JARVIS-DFT entries for bulk modulus and bandgap to numerical precision, attributing degradation to functional bias, not agent malfunction.
• Tool-augmented mean absolute error (MAE) on memorization-resistant test sets was below 0.005 eV, compared to 1.25 to 1.86 eV without tools.
• Demonstrated autonomous multi-step workflows, including 10-operation defect-engineering pipelines.

Conclusions:

• Tool access is indispensable for accurate predictions in agentic AI systems when parametric knowledge is unavailable.
• AGAPI platform effectively separates agent pipeline fidelity from inherited functional bias.
• The study confirms the critical role of integrated tools in advancing scientific discovery through agentic AI.