How open source projects need to adapt to the AI coding era

A Carnegie Mellon study found that open source projects using AI coding tools like Cursor saw a 281% spike in lines of code in the first month, but within two months velocity gains dissipated while static analysis warnings rose 30% and code complexity rose 41% — and those quality metrics didn't recover. The accumulated technical debt itself slows future development, with a 100% complexity increase leading to roughly a 65% drop in future velocity. The proposed solution draws on open source's existing contribution tooling tradition: just as CONTRIBUTING.md and PR templates shaped human contributors, project-specific agent instruction files (e.g., AGENTS.md, .cursorrules, CLAUDE.md) can encode naming conventions, architectural patterns, and quality standards to guide AI agents before code is committed. A second study on agent skills reinforces that focused, human-authored, project-specific instructions outperform generic or self-generated ones. The recommendation is to treat these instruction files as first-class project artifacts, versioned alongside the codebase and used symmetrically by both contributors and reviewers.