Your team runs a Cloud Composer (Airflow) environment that orchestrates roughly 40 daily DAGs. During the nightly peak window, dozens of DAG runs are triggered at the same time, but you notice that many tasks sit in the 'scheduled' or 'queued' state for long periods and never all run in parallel, causing SLA misses. Worker CPU and memory are consistently under 40% utilization, and the underlying GKE nodes have spare capacity. Which action most directly addresses the bottleneck?