When Karpenter is installed in a working cluster, it observes the aggregate resource requests of unscheduled pods and makes decisions to launch new nodes and terminate them to reduce scheduling latencies and infrastructure costs.