Compressed Agents.md > Agent Skills (vercel.com)

A recent exploration into enhancing coding agents' performance using Next.js 16 APIs revealed that embedding a compressed documentation index within AGENTS.md significantly outperformed traditional skills-based approaches. While the skills, designed to package domain knowledge for agent use, achieved a maximum pass rate of 79%, the embedded 8KB docs index yielded a perfect 100% pass rate during eval tests. This surprising outcome highlights the issue of agents not reliably invoking available skills, even with explicit instructions, as they might miss crucial information due to the need for decision-making. The key takeaway is that passive context, such as the AGENTS.md approach, eliminates the decision point that typically hinders skill usage, ensuring that relevant documentation is always available without the need for the agent to actively seek it. Compression techniques reduced the total context to 8KB, proving that frameworks can enhance agent performance efficiently while maintaining accuracy in code generation. This research urges framework authors to consider providing AGENTS.md snippets for their projects, as passive retrieval methods currently offer broader benefits than skills-related retrieval for general tasks, although both strategies can complement each other.