Encoding analytical workflows as agent skills for reproducible petroleomic interpretation in organic geochemistry

Ultrahigh-resolution mass spectrometry can provide thousands of molecular assignments for crude oils, but converting these data into geochemical information still relies on laboratory-specific sequences of filtering, compositional indices, visualization, and expert decisions. These procedures are often scattered across scripts, notebooks, and tacit knowledge, making them difficult to reproduce, audit, and transfer. We introduce a framework for encoding analytical procedures as reusable skills for language-model agents. The agent interprets the task and selects or coordinates skills, whereas all numerical operations are executed by deterministic, versioned code with explicit parameters and provenance. Rather than a single monolithic pipeline, the proposed framework uses a composable architecture spanning instrument access, spectral processing, compositional analysis, and chemical interpretation. As an initial implementation and proof of concept in organic geochemistry, a petroleomicscomposition skill was applied to 220 FT-ICR MS molecular-assignment tables from 110 Brazilian crude oils spanning 10 basins, comprising 1,085,391 assigned peaks. Processing six contrasting oils generated 38 figures and 49 spreadsheets in 14.8 s, while all 220 tables were processed for geochemical indices in 113.5 s. More importantly, the automated outputs recovered geochemically meaningful trends. In ESI(−), the acyclic/cyclic ratio of O2 carboxylic acids decreased with biodegradation, from a median of 1.23 in undegraded oils to 0.38 in slightly biodegraded oils (Spearman ρ=-0.701, p=9.0e-16, n=98), reproducing the expected shift toward relatively greater cyclic O2 contribution with biodegradation. In APPI(+), the thermal evolution index derived from DBE-carbon-number distributions was positively associated with thermal maturity after restricting the analysis to basins represented by at least five oils (Spearman ρ=0.412, p=3.0e-5, n=96). Within the single basin containing both lacustrine and marine oils, N2- class compounds occurred in all nine lacustrine samples but only one of nine marine samples, revealing a marked compositional contrast consistent with previously reported enrichment of nitrogen-containing species in lacustrine oils. Agent routing was evaluated with 38 queries repeated in triplicate against five decoy skills, yielding 71.1% accuracy, 57.7% recall, and 100% specificity. These results show that analytical know-how can be transformed into executable, auditable, and reusable agent skills, providing a path toward reproducible AIassisted workflows in organic geochemistry.