AI coding／軟體工程研究
RAMP Grades In-Repository AI Configuration: Complexity Growth Nearly Doubles in Unconfigured Projects After Coding-Agent Adoption
Researchers used 441 enterprise repositories to develop the four-level Repository AI Maturity Profile, then reanalyzed 509 open-source projects that had adopted coding agents. Repositories with shared rules or agent configurations showed better quality metrics, but the observational study cannot establish that the configuration files themselves caused the improvements.
Researchers have proposed the Repository AI Maturity Profile (RAMP), which measures how projects manage coding agents by scanning AI tool configurations committed to version control rather than relying on surveys. Its four levels are: L1, with no configuration at all; L2, with behavioral rules, architecture documentation, or coding standards; L3, which adds named agents, commands, and Skills; and L4, which defines multi-agent dependencies, handoffs, and execution logs.
The classifier covers 12 tools, including Claude Code, Cursor, GitHub Copilot, Codex, Gemini CLI, and OpenHands. It first uses 43 sets of path patterns to locate files, then maps them to nine semantic categories using path and content embeddings. The development sample comprised 441 private repositories from 27 organizations and 1,046 validated AI artifacts. Independent manual annotation reproduced 34 of the 35 repository-level ratings in the held-out sample. The results classified 66.7% as L1, 24.7% as L2, and 8.6% as L3; no L4 repositories were observed in the private sample. Moreover, 73.8% of configuration files were never modified after their initial commit.
The team then applied the frozen classifier to an existing dataset on coding-agent adoption. Among 509 traceable repositories, commit volume in the agent-first subsample increased by roughly 28% to 38% across all maturity levels, but quality outcomes differed: cognitive complexity rose by 53% for L1 repositories, compared with 27% for repositories at L2 or above. The increase in static-analysis warnings also differed by a factor of about 1.7. These findings suggest that `AGENTS.md`, `CLAUDE.md`, `.cursor/` rules, and reusable skills may be more than operational conveniences—they may also serve as measurable engineering control surfaces.
Engineering teams can begin by bringing agent rules into code review, assigning owners and update cadences, and tracking complexity, warnings, and rework rates. RAMP, however, shows correlation only: teams with stronger management discipline may already be more likely both to maintain configuration files and to produce better code. The private repositories cannot be disclosed individually, and L4 was absent from the development data. Prospective or randomized experiments with configurations fixed before agent adoption will be needed to determine which rules actually reduce technical debt.