In stateful agent harnesses, a procedural practice is an executable, experience-derived intervention designed to bridge the procedural-memory gap across independent executions. Rather than preserving diagnostic lessons as informal postmortem notes that an agent must rediscover or infer, a practice formalizes the lesson as an inspectable, versioned artifact—such as a state-local prompt instruction, an explicit constraint, a transition check, or a dynamic verification action triggered when visible signals indicate elevated failure risk. By encoding operational lessons into the shared runbook, procedural memory persists across runs and can be enacted mechanically by the runtime.