Methodological harnesses exist in the wild — mostly as one artifact, not a system
Sep 25 Díaz, Pérez & Gil-Borrás: 5,435 agentic repos; reweighted any-mechanism prevalence 21.7% (65.3% in >100★); mean breadth 1.15/5 even in popular stratum; rule files dominate, persistent knowledge ~1.9%.
Díaz, Pérez, and Gil-Borrás ask the precondition question most “software factory” pitches skip: do the team-level harness mechanisms people write about actually show up in repositories with agentic PRs? Using AIDev (116,211 repos), they clone and mine 5,435 stratified by visibility, operationalize seven of eight mechanisms from their prior methodological-harness framework (N-version/parallel agents leave no durable Git trace), and qualitatively code 285 rule files from 150 repos.
The receipts are sobering for continuous factories. Reweighted population prevalence of any mechanism is 21.7%; the popular (>100★) census hits 65.3%. Mean breadth among five common mechanisms is only 1.15 even in that popular stratum — 2.6% combine four or five; outside it, almost none do. What spreads is an artifact (rule/context files: H1) plus pre-agent governance (PR templates, CODEOWNERS), not the integrated method: persistent shared knowledge is ~1.9% at the top and near-absent elsewhere; executable specs stay rare. Among repos that already have a rule file, guiding the agent and stating norms are near-universal; persistent knowledge, executable specs, prescribed consultation, and graduated autonomy appear only in a minority. Temporal coverage bites: 17.7% of popular-tier agentic PRs look harnessed at clone time but were not when opened — evaluating HEAD after the fact invents overnight governance that wasn’t there for the merge.
They deliberately measure presence, not efficacy (Faros/DORA/METR tensions stay as motivation). Structured consultation is real in comments but their lexical detector’s recall (~0.49) forced exclusion from the quantitative table — instrument failure, not mechanism absence.
Why it matters for code review
FOMO for always-on agents is cheap; an integrated methodological harness is not. Review the system, not the CLAUDE.md selfie: if overnight merge trains lack persistent decisions, executable acceptance evidence, and autonomy boundaries, you have a rule file riding CI — not a factory. Pin breadth and co-occurrence the way this paper does before declaring the team “harnessed.”