Why AI Agents Ignore Your Documentation and What Actually Works

A developer building an AI agent framework discovered that placing critical knowledge in an MCP resource made it effectively invisible, because agents rarely choose to fetch passive information on their own. The core insight is that agent knowledge delivery depends on trigger conditions, not content quality — each storage method fires under different circumstances and carries different context costs. Always-on files like AGENTS.md load every turn and should only hold routing logic and universal guardrails, while tool-based and skill-based knowledge loads only when intent or situational matching occurs. Workflow instructions placed in passive resources or docs often go unread because nothing directs the agent toward them at the right moment. The developer concludes that most agent knowledge failures stem from correct content placed in the wrong trigger channel, not from poor writing or incomplete information.