This work investigates the use of large language model (LLM) agents to discover compact, high‑accuracy exchange‑correlation (XC) functionals within a GPU‑based evaluation loop. By linking frontier LLM agents (Opus‑5.5) to a self‑consistent workflow through the Model Context Protocol, we enable training‑only fitting, persistent memory, and both self‑consistent and external evaluation. Compared with Codex (GPT‑6‑Astra) agents, Opus‑5.5 agents consume roughly 40‑fold fewer tokens during the screening stage and identify a candidate that improves both validation and held‑out errors. The sole candidate advancing to self‑consistent testing is the 13‑parameter functional O159, which multiplies the same‑spin correlation term by a smooth density window.
Refitting O159 shows that the window scales the entire same‑spin factor, reducing its peak by $15\%\sim30\%$ and primarily affecting reaction energies in near‑iso‑orbital regions. Against a no‑window control, O159 lowers heavy‑atom total‑energy errors on the GSCDB137 benchmark by $29\%$. In contrast, SyFES evolution with about $400$ fits each fails to improve upon its GAS22 seed, whereas the agents achieve a validation error reduction of $0.10\sim0.13$ \kcal by adding an explicit density scale. Self‑consistent O159 reduces the held‑out MGCDB84 test error of $\omega$B97M‑V by $11.7\%$, and its mean normalized error ratio across 36 GSCDB137 datasets is $1.139$, comparable to $\omega$B97M‑V ($1.144$) and better than GAS22 ($1.190$).
Review: Frontier LLM agents can dramatically shrink the search space while preserving physical interpretability, yielding XC functionals that outperform traditional evolutionary methods and survive rigorous self‑consistent, physics‑based validation, highlighting the strong synergy between AI and quantum chemistry.