Your monitoring can tell you something broke. This book is about what acts next. Threshold alerts and hand-written runbooks were built for infrastructure an engineer could hold in their head. That assumption broke somewhere between the first service mesh and the first GPU fleet, and the gap is now filled by paging a human at two in the morning. This handbook sets out ten worked approaches to closing it with agents that reason rather than react. Compositional verification for alerting with provable safety bounds. Runbooks synthesized under symbolic constraints. Counterfactual chaos experiments for serverless systems that cannot hold a control group. Causal diagnosis ranked by expected utility, online causal masking across logs, traces, and metrics, and hierarchical abstraction over causal graphs that will not sit still. It closes on the operator, and on how an agent earns the trust it is given. Every chapter carries a formal statement, an experimental evaluation, and its ablations. Written by two engineers who run this infrastructure for a living, for the researcher and for the person on call.