Financial Agent Skill Engineering: A Grounded Theory Study of 80 Production Skills Across Eight Domains

Abstract

Agent Skills are increasingly deployed as production assets in financial institutions, yet build-time package architecture remains scarce. We present a Grounded Theory study of 80 financial Agent Skills (65 business-built + 15 FinClaw rebuilds) across eight sub-domains. Primary contributions are seven architecture skeletons (A–G) and a friction-driven standards account (90 events, 64 backfilled). Skeletons A–E and G reach coding stop after 5–12 samples each. Code-Gen F stops after six samples that span six sub-patterns; selective skeleton assignment on a stratified sample of n=45 packages shows substantial inter-coder agreement (Cohen's κ=0.69), with residual softness at the F/B boundary. Under a Factory packaging rubric, all 15 rebuilds score higher than FinClaw sources (mean Δ+41.8; mostly on engineering-leaning dimensions). Content ratings and six internal cases are secondary and exploratory. Rubrics, protocols, and a de-identified friction taxonomy are released; raw sources remain withheld. This record is a preprint of a manuscript submitted to Empirical Software Engineering (EMSE). Accompanying files: (1) open replication package (scoring rubrics, IRR/Content protocols and sheets, GT selective-coding materials without KEY, tables, CANONICAL numbers); (2) a stand-alone de-identified friction taxonomy of 90 events across 8 financial sub-domains spanning framework versions v1.0→v1.6 (64/90 backfilled; client identifiers and proprietary thresholds removed). Commercial skill sources and the raw friction index with business detail remain withheld. Version 1 (2026-08-05): preprint matching the EMSE submission package (svjour3 non-referee layout). Open zip SHA256: DE865EA61C238AB03290E8C410252A96F4D9440FCDCE2B368AE6A110D5EF72CB. GT coding stimuli (45 anonymized packs) withheld for confidentiality; GT verification chain (codebook, filled coder sheets, kappa summary) is included.