A few pages of committed markdown, half the quality damage from coding agents
Across 509 open-source repositories, coding agents made every team faster. But in repositories with no committed AI configuration, code complexity grew about twice as fast as in repositories that committed at least a rules file and coding standards. The authors turn this observation into RAMP, a four-level maturity scale you can compute from any repository's files.
The problem
Coding agents raise development speed, and they also raise technical debt. Prior studies report both effects, but only as averages across all adopters. Some teams report large, durable gains from agentic workflows. Other teams describe agent output as a burden on reviewers. The averages cannot explain this split, because no study measured how each team set up its agents.
The authors of this paper start from one observation: part of that setup is committed to version control. A CLAUDE.md or .cursorrules file, a coding-standards document, an agent definition with allowed tools — these files are objective, timestamped, and readable at scale. They are a measurable record of how a team works with AI.
RAMP: a maturity ladder built from committed files
The paper defines RAMP (Repository AI Maturity Profile), a four-level scale. The committed AI files are called artifacts, and a repository sits at the highest level for which it has at least one:
Level 1 — Unconfigured. No AI-related files. Each agent session starts with no project knowledge.
Level 2 — Grounded prompting. Context files: behavioral rules, tool configuration, architecture documents, coding standards. The agent now knows how the project works.
Level 3 — Agent-augmented. Capability files: named agents with roles and tool restrictions, reusable commands, skill guides. The agent now knows how to do specific tasks the team's way.
Level 4 — Orchestration. Coordination files: multi-agent workflows with phases and handoffs, and session logs from agent runs.
A classifier assigns these levels automatically. It matches known file patterns for 12 coding tools, then classifies the remaining files by meaning: it embeds each file's path and content as vectors and compares them with nine category templates. The authors validated the output against human annotators, who reproduced the classifier's repository-level label for 34 of 35 held-out repositories. Statistical tests confirm the levels form a cumulative ladder: a Level 3 repository almost always also has Level 2 files, and repositories combine categories in only 25 distinct ways out of hundreds possible.
How teams actually adopt configuration
Study 1 applies the classifier to 441 private repositories from 27 companies. Most repositories sit at the bottom of the scale: 67% have no committed AI configuration at all, 25% reach Level 2, 9% reach Level 3, and none reach Level 4.
The timing pattern is clear. Teams take a median of 633 days of observed activity before they commit their first AI artifact. Once a team commits one, the next level takes only about 154 days. Movement is forward-only: the authors found zero reversals. And the files rarely change afterward: 74% of artifacts are committed once and never modified. Whatever you commit in the first weeks of agent use will likely govern your agents for the life of the project.
Results: speed for everyone, debt for the unconfigured
Study 2 uses an existing dataset of 509 open-source repositories that adopted coding agents, and splits it by RAMP level: 236 at Level 1, 273 at Level 2 or above. The statistical method is a difference-in-differences design: it compares each repository's trajectory before and after agent adoption against matched repositories that did not adopt. The headline numbers below come from the subset in which an agent was the repository's first AI tool; this holds prior AI exposure constant.
Speed rose in both groups: 38% more commits at Level 1, 28% more at Level 2+. Level 2+ repositories added more code (+69% lines versus +48%). Configuration does not slow agents down.
Quality moved in the opposite direction. After agent adoption, cognitive complexity — a static measure of how hard code is to follow — rose 53% in Level 1 repositories against 27% at Level 2+, a 2.0× gap. Static-analysis warnings rose 24% against 14%, a 1.7× gap. A four-level breakdown shows the same order at every step: the complexity increase shrinks at each level, and at Levels 3 and 4 it is statistically indistinguishable from zero.
The limits matter, and the authors state them clearly. This is an observational split, not a randomized experiment. Teams that commit configuration may simply have more engineering discipline, or use more capable models; the study design cannot separate those explanations from a direct effect of the configuration files. Most Level 2+ repositories committed their configuration alongside or after agent adoption, not before. The warnings result rests on weaker evidence than the complexity result, because Level 1 repositories already showed elevated warnings before adoption. And even configured repositories still saw a significant 27% complexity increase — configuration reduced the damage; it did not remove it. The authors label the results hypothesis-generating: they suggest a cause, but they do not prove one.
What you can do with it
The actionable gap is the first one: from nothing to a few pages of committed markdown. A rules file plus a coding-standards document is cheap to write, and its presence is associated with roughly half the quality degradation. Higher levels add only smaller reductions. If your team is adopting agents, commit those two files in the first months — the files rarely change later, so the initial configuration is the one that will persist.
The classifier is also useful across a whole organization. It needs no questionnaire and no cooperation from the team it measures: point it at a hundred repositories and see which ones face an agent rollout with no committed configuration. The pipeline runs locally on embeddings, so private code never leaves your infrastructure; the authors note that an LLM judge with the same category definitions performs about as well if you can send files out. The replication package is public.