On the Maintenance and Co-evolution of Agent Plugins: An Empirical Study of Claude Code Plugin Marketplaces
A study of 8,351 Claude Code plugins finds instruction files and scripts often require coordinated maintenance, treating skills as evolving software rather than static prompts.
Researchers analyzed **8,351 plugins** and **77,773 commits** across Claude Code marketplaces. Software-engineering plugins made up 61.3%, Claude co-authored 34.9% of commits, and plugin-touching activity grew 8.8× over six months.
Treat agent plugins as maintained software artifacts. In skills directories, instruction files and implementation scripts changed together above chance, and **78% of those co-changes** were judged functionally coupled; reviews and tests should cover both sides of that boundary.
Researchers analyzed **8,351 plugins** and **77,773 commits** across Claude Code marketplaces. Software-engineering plugins made up 61.3%, Claude co-authored 34.9% of commits, and plugin-touching activity grew 8.8× over six months. Treat agent plugins as maintained software artifacts. In skills directories, instruction files and implementation scripts changed together above chance, and **78% of those co-changes** were judged functionally coupled; reviews and tests should cover both sides of that boundary. The study examines repositories associated with one agent ecosystem after its October 2025 launch. Commit classification and observed co-change indicate maintenance patterns, but do not by themselves measure plugin quality, reliability, or user adoption.
The study supplies ecosystem-scale evidence that plugins and skills behave like evolving software, not static prompt files. Its observed coupling between instructions and scripts strengthens prior calls for shared, governed agent workflows and makes joint tests and reviews an implementation consequence. It does not, however, validate the quality, adoption, or portability claims of any particular skill catalog.