Why AI Agents Need Runbooks That Document Wrong Answers, Not Just Right Ones

A developer running a fleet of Claude Code AI agents in tmux panes discovered that one agent had introduced a critical bug while attempting to fix a problem that never existed — changing the Enter key submission to a backslash-Enter, which silently prevented all prompts from sending. The root cause was the agent's lack of episodic memory: each new context window starts fresh and will independently re-derive the same plausible but incorrect fix every single time. This led the developer to build runbooks that explicitly document refuted hypotheses, not just correct procedures, since AI agents cannot learn from past mistakes the way human teams do through shared experience. A similar issue arose in a healthcare billing system, where a quality gate anchored outputs to a weaker AI model scoring only 38% accuracy against golden answers, compared to 78% for the newer system it was meant to validate. The key lesson is that for AI-driven operations, documentation must also serve as a graveyard for ideas that seemed reasonable but were proven wrong, to prevent future sessions from confidently rebuilding broken solutions.