Summary
Claude Code’s subagent system enables a main orchestrator session to spawn isolated child sessions that run in parallel, each independently configurable with its own model (Haiku, Sonnet, or Opus), persona, tool permissions, and project memory. In this detailed walkthrough, Nate Herk explains the core rationale — preserving the main context window from pollution during research or file-heavy tasks, routing expensive work to cheaper models, and enabling genuine parallelism across specialized agents — and then shows how to build and configure them in practice.
The video covers the distinction between built-in subagents (like Claude Code’s automatic research agent, which many users encounter without realizing it) and custom agents defined via Markdown files. Herk walks through creating a new project-scoped agent using the `/agents` UI, including how to describe the agent’s purpose, set tool access to read-only or full permissions, configure individual tools like Bash and file creation, select the underlying model, and enable specific MCP servers. He also explains the critical scope decision: agents stored globally in `~/.claude/agents/` are available across all projects, while those in `.claude/agents/` within a project directory are scoped locally.
A practical opening demo shows five subagents running concurrently with distinct personas — including a retired elementary school teacher and a Fortune 500 COO — each reviewing the same document from a different perspective, with the main session synthesizing their outputs. For developers using Claude Code at scale, this video is one of the more thorough practical references available for subagent architecture.
📺 Source: Nate Herk | AI Automation · Published June 09, 2026
🏷️ Format: Tutorial Demo
