Researchers from the Federal University of Minas Gerais have identified significant inefficiencies in how AI agents manage instruction files, leading to unnecessary token consumption and degraded output quality. By analyzing 532,000 files such as CLAUDE.md and AGENTS.md across 100 open-source projects, the study found that many developers are stuffing agents with redundant rules and global skills that should be modular.
These systemic issues stem from a lack of optimization in agent configuration. Key findings from the research include:
- Redundant Validation: Many files contain coding rules that are already automatically handled by standard linting tools, forcing the AI to process known constraints repeatedly.
- Information Overload: Detailed project contexts and specialized "skills" are being loaded globally rather than on-demand, which fills the context window with noise.
- Token Waste: The excessive length of these files directly correlates to higher operational costs and slower response times for developers using AI-assisted workflows.
Optimizing these files by removing repetitive constraints and implementing lazy-loading for specific agent capabilities can significantly improve efficiency in open-source AI development environments.
