Subagents separate context, not usage limits
Claude Code documents focused delegation inside a session, with requests still counted against shared usage.
Delegate a bounded investigation when you need its result without keeping every search in your main conversation. First define what the subagent must return; its requests still use the conversation's shared allowance. The Claude Code subagent documentation describes existing assistants with separate context, instructions, tools and permissions, rather than a new release.
A subagent works separately and returns its result to the main conversation. That can help when a task involves extensive searches or file reading whose intermediate material does not need to remain in the central context. A bounded assignment also makes the expected return clearer than an open-ended request to investigate everything.
Subagent requests count toward the same usage limits as the main conversation. Separate context windows do not, by themselves, establish lower cost, faster completion or better results. Those outcomes depend on the assignment and on how much information must return to the coordinating assistant.
- Specify the output and include the information needed to produce it.
- Coordinate edits before assigning work on shared code. Check whether you need a subagent within this session, an independent session or an agent team; these are different arrangements. Confirm that the returned result contains what you need for the next decision.