Empirically maps structure and co-evolution of AI coding-agent plugin marketplaces, highlighting delivery via natural-language instructions, scripts, and configs rather than source code.
On the Maintenance and Co-evolution of Agent Plugins is a study analyzing 1,926 repositories, 8,351 plugins, and 77,773 commits from the Claude Code plugin marketplace. It reveals that these plugins are AI-native artifacts delivering functionality through natural-language instruction files, scripts, and configuration files rather than traditional source code.
The study highlights several key empirical findings: Co-evolution Dynamics: While most components evolve independently, Script–Markdown coupling within skills represents a new maintenance dependency where natural-language instructions and implementation scripts change together at above-chance rates. Maintenance Patterns: Plugin development is predominantly feature-driven, with feature commits occurring at more than twice the rate of conventional open-source software. AI Contribution: AI coding agents co-author 34.9% of all commits, indicating a significant shift in human-AI collaboration dynamics. Artifact Nature: Natural-language files (primarily Markdown) serve as the primary distributed artifact, requiring new empirical baselines for software maintenance and co-evolution analysis.
This paper presents an empirical study of plugin ecosystems for Claude Code, focusing on how plugins are structured, distributed, maintained, and allowed to co-evolve within AI coding-agent marketplaces. Rather than treating plugins as conventional software packages, the study maps a more heterogeneous artifact landscape: many plugins are delivered primarily through natural-language instructions, auxiliary scripts, and configuration files, with comparatively little traditional source code. This framing positions plugin marketplaces as hybrid ecosystems where prompts, executable snippets, and declarative settings all function as core components of the agent’s behavior.
A key contribution is the paper’s attention to maintenance and co-evolution in this setting. Plugin updates may involve prompt refinement, script adaptation, configuration changes, or alignment with evolving agent capabilities and task conventions, rather than only bug fixes or feature additions in compiled code. The study therefore highlights how plugin quality and compatibility depend on interactions among the host agent, the plugin’s linguistic and procedural components, marketplace norms, and user workflows. It suggests that versioning, dependency tracking, and compatibility management in agent plugin ecosystems require new abstractions that account for natural-language behavior and runtime configuration as first-class elements.
The work matters because it documents an emerging software ecosystem with implications for agent reliability, security, governance, and lifecycle management. As coding agents become more capable and plugin marketplaces grow, understanding how plugins are maintained and how they co-evolve can inform the design of better tooling for evaluation, reproducibility, trust, and dependency tracking. More broadly, the paper points to a shift in software engineering practice: in agent-centric systems, maintainability is not only a property of code, but also of the instructions, scripts, and configurations that shape agent behavior.