The problem: operators were tab-hopping across ten Brain pages to piece together system state. Cognitive load was the bottleneck, not information.
The pattern: one server function aggregates read-only snapshots from every Brain in parallel, produces a single normalized payload, and passes it to the model as grounded context. The UI shows summary + prioritised actions + ask-anything.
Rule: any new admin surface must integrate into Kernel Copilot's fan-out before adding its own top-level nav entry. Consolidation over accumulation.