Subagents

A single agent works linearly and wastes time on parallelizable subtasks. Subagents let them start together.

How it works

The main agent uses spawn_agent to hand a subtask to a fresh agent instance, carries on with something else, then consolidates the results.

When it helps

• The task splits into independent pieces (research three libraries at once)
• You want several separate outputs (a summary per module)
• A broad search where fan-out meaningfully cuts the wait

Keep in mind

Each subagent costs its own tokens and context. They work within their own boundaries, and the main agent integrates their output rather than handing you a pile of fragments.