How Claude Code Uses Subagents for Parallel Work

Anthropic published a deep guide on Claude Code's subagent system - isolated Claude instances that spin up with fresh context windows, run tasks in parallel, and return results without inheriting the main conversation's assumptions. Three built-in types ship by default: general-purpose, plan, and explore agents, each with different permission scopes. Custom subagents can be defined as markdown files in .claude/agents/ with their own system prompts, tool access, and model overrides. The practical gain is parallelizing research across dozens of files or running unbiased code review from a clean context. This release lands alongside today's Managed Agents API announcement, and together they paint a clearer picture of Anthropic positioning Claude as a multi-agent orchestration platform rather than a single-thread assistant.