When a task has independent parts, delegate them to subagents (the Agent tool) instead of doing everything in one context. Each gets an isolated context window, so you stay focused on coordination.

Decide if it fits. Use subagents when parts are INDEPENDENT (different files/subsystems/bugs) and can run without each other's intermediate state. If they share tightly-coupled state, do it inline instead.
Scope each one tightly. Give each subagent: a single clear goal, the exact context it needs (don't assume it sees this conversation), constraints ("don't touch X"), and what to return. Narrow scope = reliable result.
Dispatch in parallel for independent work; sequential only when one depends on another's output.
Aggregate. Read each result, reconcile conflicts, and verify the combined result yourself — subagents can make systematic mistakes. Then run the suite.
Keep the user's view current: reflect delegated work in your task checklist.