Software repositories became more active after coding-agent adoption, but the apparent quality cost was smaller where teams had more structured AI practices committed to the repository, according to a new preprint. In repositories that used an agent first, cognitive complexity rose 52.7% in the least mature group, compared with 26.7% in repositories at Level 2 or above. Static-analysis warnings increased 24.1% versus 14.0%, a 1.7-fold contrast.
The same repositories also recorded more commits after adoption: up 37.6% at Level 1 and 27.5% at Level 2 or above. Lines of code added rose 48.1% in the lower-maturity group and 68.7% in the higher-maturity group. The pattern suggests that greater activity and different quality trajectories can occur together, but it does not show that committed rules or standards caused the difference.
A maturity scale built from committed files
The study introduces RAMP, a four-level measure of repository-level AI-practice maturity. It looks for committed, version-controlled artifacts such as configuration and guidance files, using filename patterns and semantic classification across 12 AI coding tools. The model treats the levels as cumulative, with higher levels representing additional kinds of committed practice.
In the development set of 441 corporate repositories from 27 organizations, 294 repositories, or 66.7%, were at Level 1. Another 109, or 24.7%, were at Level 2, while 38, or 8.6%, were at Level 3. No repository reached Level 4.
Several checks supported the proposed ordering. A scalogram test, which asks whether items tend to appear in a cumulative sequence, produced a coefficient of reproducibility of 0.997 and a coefficient of scalability of 0.983, above the stated thresholds of 0.90 and 0.60. In a separate human check, the classifier matched the assigned maturity level in 34 of 35 repositories, or 97.1%. Agreement at the individual-file level was 81.7%, rising to 93.6% for references backed by a majority of annotators.
Most repositories stopped after the first step
The reconstructed adoption history points to a front-loaded process. Among the repositories examined over time, 95.9% entered Level 2 first, and 83.1% added artifacts at only one level. A smaller 14.9% moved from Level 2 to Level 3. The study recorded no reversals, while the median time from Level 2 to Level 3 was 154 days.
Most configurations also appeared to be set and then left alone: 73.8% were classified as set-and-forget. The result adds to the picture of an adoption process that was usually front-loaded and forward-moving.
The comparison followed agent-first repositories
For the outcome analysis, the researchers reanalysed a prior panel of repositories, separating agent-first repositories by maturity level. The analysis began with 518 treated repositories; after nine could not be traced, 509 remained, including 236 Level 1 repositories and 273 at Level 2 or above. The study used matched controls from the original panel.
The researchers estimated changes in monthly commits and lines added, alongside SonarQube measures of cognitive complexity, static-analysis warnings, duplicated-line density and comment-line density. They used a staggered difference-in-differences approach, a method that compares changes over time between treated repositories and comparable controls, with repository-clustered standard errors.
Additional checks kept the main quality pattern for the lower maturity levels. Cognitive complexity rose 54.4% at Level 1 and 39.8% at Level 2, while the estimates at Levels 3 and 4 were indistinguishable from zero. Level 4 was an exception for duplicated-line density, which rose 31.4%.
An association, not a causal verdict
The authors describe the findings as hypothesis-generating rather than causal. Warnings already showed elevated pre-treatment coefficients and were more weakly identified, making it difficult to rule out the possibility that some differences preceded adoption.
The outcome analysis used SonarQube measures such as cognitive complexity and warnings rather than a single quality score. Robustness checks preserved the complexity pattern at Levels 1 and 2, while Level 4 was an exception for duplicated-line density.
The paper says a replication package contains the analysis pipeline, tool patterns and notebooks. The outcome data for the second study come from a public prior dataset, while the identities of the corporate repositories used to develop RAMP are restricted.
Paper data and sources
Original title: A Few Pages of Markdown: Committed AI Configuration and Lower Quality Cost after Coding-Agent Adoption
Authors: Yegor Denisov-Blanch, Shyam Agarwal, Pavel Azaletskiy et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text