Repos that commit an AI config file take half the complexity damage. 73.8% commit it once and never touch it again.

Stanford's Denisov-Blanch group built a maturity model for AI adoption scored entirely from artifacts already in version control, applied it to 441 repositories, and found something I've been assuming without evidence.

RAMP is a four-level model derived only from committed AI configuration files. No surveys, no self-reporting, just what's in the repo. Independent human annotation reproduced RAMP's labels on 97% of a held-out sample, which is a stronger validation than most tooling papers bother with (arXiv 2608.25241).

The volume finding is uniform and boring: agents raise commit volume 28 to 38% at every maturity level. Config or no config, mature or not, more commits.

The quality finding splits hard. Among agent-first repos, those with no committed AI configuration show roughly twice the increase in cognitive complexity, +53% against +27%, and 1.7x the increase in static-analysis warnings. Same productivity lift, double the mess.

Then the number that made me close the tab and open my own repos: 73.8% of these configuration artifacts are committed once and never modified again. Someone runs claude init or drops in an AGENTS.md, gets the file, commits it, and never edits it again as the codebase changes underneath it.

I'm guilty of a version of this. My CLAUDE.md files get written at project start, get one or two panicked edits after an agent does something dumb, and then sit. The lesson I'm taking from this paper is that the config file is a living document with the same maintenance obligation as a test suite, and I've been treating it like a README.

The authors are careful, and I'll be careful with them. This is observational and they flag it as hypothesis-generating. Teams that commit and maintain agent configuration are plausibly teams that were already disciplined about complexity, and the config file is a marker rather than a cause. Nobody has run the controlled version.

But the intervention costs nothing. Commit the file. Then put a recurring item on your calendar to actually edit it. That's the whole action, and if the causal story is even half true, it's the highest-leverage 20 minutes in your month.

This connects to something Warp published about their internal setup: a two-skill loop where a base skill holds domain instructions and an outer "improver" skill runs on a schedule, pulls accumulated human feedback, compares the agent's suggestions against what humans actually did, and opens PRs editing the base skill (Anthropic). Their operating rules are specific and I've started copying them: write principles rather than rules, always include the rationale so the agent generalizes, keep skills small and reference external files, and weight one senior engineer's detailed PR comment above high-volume cursory signal. Warp reports 10M+ Claude Code sessions internally, 400K+ per week, and calls the improver skill the reusable half.

Automating the edit is the answer to the 73.8%. Warp built it. The rest of us are still editing by hand, when we edit at all.