Split work across subagents so each has a focused context, and combine their results deliberately. Use when a task is too large for one context or has independent parallel parts.
