A Few Pages of Markdown: Committed AI Configuration and Lower Quality Cost after Coding-Agent Adoption
Yegor Denisov-Blanch, Shyam Agarwal, Pavel Azaletskiy, Hao He, Rylan Schaeffer, Brando Miranda, Bogdan Vasilescu, Sanmi Koyejo
cs.SE, cs.AI
2026-08-26
RAMP scores committed AI config on four levels. Agents raise commits 28-38% at every level; agent-first repos with no config see +52.7% complexity versus +26.7%.
Coding agents make some teams durably faster and leave others reviewing slop. Existing studies report averages over adopters: short-run velocity gains, persistent rises in static-analysis warnings and complexity. Averages hide the spread across teams. What is missing is a practice measure that can be observed at repository scale and aligned with code-quality outcomes.
Industry AI maturity models from the past year mostly score organizations with surveys. Quality is a property of a codebase, so practice has to be measured at the same unit. Many adoption events have already happened, so the measurement has to come from what teams left in version control.
RAMP (Repository AI Maturity Profile) is a four-level cumulative model. A repository's level is the highest artifact class it contains: L1 has no committed AI configuration; L2 has behavioral rules, tool settings, architecture notes, or coding standards; L3 has named agents, reusable commands, or skills; L4 has multi-agent orchestration. A pipeline covers path patterns for 12 tools, then classifies path tokens and file contents with nomic-embed-text-v1.5 against nine category templates.
The classifier was developed on 441 private repositories from 27 companies. Guttman CR is 0.997. Three annotators blindly labeled 195 artifacts across 35 repositories from the Study 2 corpus; repository-level labels match for 34/35 (97.1%). Study 1 tracks transitions among 196 adopters over about 14 months. Study 2 applies the frozen classifier to Agarwal et al.'s open-source agent-adoption panel, stratifies 509 traceable treated repositories into L1 versus L2+, and re-estimates staggered difference-in-differences inside the agent-first stratum.
In the 441-repository development sample, 66.7% sit at L1, 24.7% at L2, 8.6% at L3, and none at L4. Rules files are near-universal: 98.2% of L2 and 84.2% of L3 repositories have one. Adoption is single-shot, forward-only, and mostly set-and-forget: 73.8% of artifacts are committed once and never modified; 95.9% of adopters enter at L2; reversals are 0%. Median time to the first artifact is at least 441 days; L2 to L3, when it happens, has a median of 154 days.
Study 2's treated set is 236 L1 and 273 L2+ (including 40 open-source L4). Velocity rises in both agent-first strata: commits +37.6% (L1) versus +27.5% (L2+), lines added +48.1% versus +68.7%. Quality diverges.
| Metric (agent-first) | L1 unconfigured | L2+ structured | Ratio |
| Cognitive complexity | +52.70% | +26.68% | 2.0× |
| Static-analysis warnings | +24.08% | +14.04% | 1.7× |
| Duplicated line density | +15.44% | +14.23% | 1.1× |
| Monthly commits | +37.56% | +27.52% | 1.4× |
He et al. reported +30% warnings and +41% complexity for Cursor adopters. L2+ sits below both baselines; L1 exceeds the complexity baseline. Pre-treatment coefficients for warnings are elevated, so that contrast is more weakly identified than complexity. The authors frame Study 2 as hypothesis-generating, not causal.
For teams that already let agents touch production code, the paper offers a cheap stratification: look at whether the repo committed behavioral rules and standards. Velocity gains appear at every level. The quality cost concentrates in repositories that commit nothing. L2+ does not eliminate the complexity increase; a significant +26.7% remains. Maturity measures presence of artifacts, not content quality or enforcement. Correlated engineering discipline or model capability may explain part of the gap. RAMP's value is that later studies can sample on how a repo configures AI, not only on whether it adopted agents.
Maturity is labeled at the end-of-panel snapshot. Only 3.8% of L2+ repositories in the agent-first stratum committed their first validated artifact before the adoption month, so reverse causality (write config after quality problems) would shrink the gap. L2+ repositories are far more starred and active; the shared control pool does not guarantee within-stratum balance. The IDE-first L1 cell has 22 repositories, so the contrast is identified almost entirely among agent-first repos. The development sample has no L4; the four-level claim leans on mapping permutation and prevalence ordering. Inter-annotator agreement is moderate (Krippendorff's α=0.572). Outcomes are SonarQube proxies, not defects.