vibe-qc: expert-directed agentic engineering for scientific software development

Generative artificial intelligence (AI) can implement scientific software, but useful development also requires executable comparisons, inspectable results and a record of scientific decisions. This case study examines the tooling that enables semi-autonomous electronic-structure software development under expert direction. VIBE-QC supplies molecular and periodic calculations; vibe-scheduler queues and manages calculations across clusters and a fleet of workstations; vibe-viewer and the Quantum Visualization Format (QVF) connect numerical outputs to portable inspection and reproducible figures. Retained development records illustrate how these tools preserve failed attempts, distinguish changes in physical conventions from implementation errors, and separate measured differences from expert-selected acceptance criteria. Publication-specific calculations were run with VIBE-QC 0.18.x and use formaldehyde to connect controlled numerical comparisons with visual inspection; diamond and lithium hydride provide periodic Perdew–Burke–Ernzerhof (PBE) examples with explicit convergence checks. The expert sets the physical problem, selects diagnostic controls and interprets the evidence; AI agents implement and execute those tasks. The resulting account identifies practical requirements for sustained AI-assisted scientific development. It establishes a traceable process within one project, without claiming autonomous scientific judgment or a measured productivity advantage.