LLMs vs. Agents as Docs Consumers (dacharycarey.com)

A recent deep dive into the consumption patterns of documentation by AI systems distinguishes between two primary types: model training and agent workflows. While large language models (LLMs) like those from Anthropic and OpenAI utilize documentation primarily for training, consuming content in bulk for inclusion in their knowledge base, coding agents such as Claude Code and GitHub Copilot interact with documentation in real-time, retrieving information to assist developers instantly. This distinction is significant as it informs how documentation should be structured and optimized for AI consumption, depending on the intended use. The implications for documentation teams are substantial; optimizing for model training emphasizes crawlability, accuracy, and clear structure, paralleling SEO best practices. In contrast, agent-focused optimization requires attention to content size, markdown availability, URL stability, and providing a structured way for agents to discover documentation—factors that can significantly affect the agent's performance. As agents gain adoption, ensuring documentation meets both consumption patterns is vital to prevent truncation errors and misinformation, ultimately improving developer experiences. Practitioners are encouraged to prioritize agent optimization and utilize tools like afdocs for automated evaluations, addressing the pressing need for clarity in an increasingly AI-integrated landscape.