One chemist directed AI agents to write a quantum-chemistry code, and molecular Hartree–Fock energies match independent programs to the nanohartree
Synopsis
vibe-qc is an open-source electronic-structure code for molecules and solids whose code was written almost entirely by AI agents under the direction of one domain scientist. It puts molecular and periodic calculations behind a common Python interface and C++17 numerical core, and ships a reusable validation set: molecular Hartree–Fock energies agree with independent programs at the nanohartree scale, and conventional CCSD(T) is compared on small molecules. The set records independent-program comparisons, published benchmarks and internal consistency separately, and keeps comparisons that miss tolerance.
Interpretation
A general-purpose electronic-structure code can now be written mainly by AI agents, with a domain scientist choosing the physical models, judging results and approving releases. Implementing and maintaining electronic-structure methods has long depended on students and postdocs writing code by hand, and programs often outlive their authors, which is why community libraries exist to reduce reimplementation. A first Hartree–Fock prototype was checked against an independent program after one day, the initial broad implementation took roughly ten weeks, and validation and revision continued afterward.
The program places molecular and periodic calculations behind one interface, covering mean-field, perturbation, coupled-cluster, active-space and excited-state methods for molecules, and Hartree–Fock and Kohn–Sham DFT for periodic systems. A C++17 core handles integral-driven Fock builds, lattice sums, derivative kernels and exchange–correlation grids, while Python handles calculation setup, file input and output, and workflow control. Integrals, functionals, crystal symmetry, effective-core-potential integrals and Fourier transforms call libint, libxc, spglib, libecpint and FFTW respectively; molecular comparisons use orca 6.1.1 and openmolcas as references, and periodic Gaussian calculations use pyscf and crystal 23.
Perspective
The demonstrated scope is small-system comparisons for molecular Hartree–Fock, Kohn–Sham DFT, MP2, CCSD(T), active-space methods and linear-response excited states, plus periodic Hartree–Fock and DFT through the GDF and BIPOLE routes; GPW/GAPW, variable-cell optimization, periodic effective-core-potential runs through the public drivers and correlated periodic methods lie outside the validated scope.
The author states that the roughly ten-week development progress describes one project and is not a controlled estimate of productivity; local correlation remains a pilot implementation whose truncation error is reported separately; in a colocated timing control on hydrogen peroxide at DLPNO-CCSD(T), vibe-qc was slower than orca 6.1.1, and the author says the implementation requires further optimization.