You can spawn multiple Claudes that work on different parts of your task simultaneously

When a task is too big for one conversation, Claude Code can spawn subagents — specialized AI assistants that handle specific subtasks in their own context windows.

Why would you want this? Context preservation. If Claude searches through 50 files to understand your API, all those file contents flood your main conversation's context window. With a subagent, that exploration happens in a separate context — only the summary comes back to you.

You can create custom subagents by dropping markdown files in .claude/agents/. Each one gets a name, description, custom system prompt, and specific tool access. For example, a "code-reviewer" agent that only has read access, or an "explorer" agent optimized for codebase research.

Claude automatically decides when to delegate based on each subagent's description. You can also force delegation with subagent_type in the Agent tool.