mirror of
https://github.com/tiennm99/ai-coding-workflow-labs.git
synced 2026-09-18 02:18:58 +00:00
tiennm99
2b505dd5e1
Add initial lesson structure for Vin Obsidian Workflows mini-lesson: - CLAUDE.md: full lesson script and teaching instructions - my-vault/: 10 interconnected Obsidian notes about Claude Code best practices - .claude/commands/start-lesson.md: slash command to launch the lesson - my-commands.md: documentation for custom commands
1.6 KiB
1.6 KiB
Subagents
Subagents are Claude Code's way of running parallel workers. Each one gets its own fresh context window — like assigning a task to a new employee who isn't overloaded with everything else you've been working on.
The key insight: subagents solve the Context Window Management problem. Instead of cramming everything into one conversation, you delegate isolated tasks to fresh agents.
The Three Types
- Explore agent — reads files, searches codebase, answers questions about code. Read-only.
- Code agent — can read AND write files, run commands. Full power.
- Custom agent — via the Task tool with specific instructions.
When to Use Them
Boris Cherny runs 5 local sessions + 5-10 web sessions in parallel. That's extreme for most people. The practical version:
- Research tasks — send a subagent to explore a codebase or read docs while you keep working
- Independent tasks — two features that don't touch the same files? Parallelize.
- Context isolation — prevent a debugging rabbit hole from polluting your main session
- Verbose operations — code reviews, file analysis, anything that generates a lot of output you don't need to see
The Pattern
Each subagent gets:
- A clear task description
- Relevant file paths or search scope
- Expected output format
- Its own fresh context window
The main session stays clean. You get a summary back.
Related
- Context Window Management — subagents are the #1 tool for managing context
- The Junior Employee Mental Model — subagents ARE the junior employees
- Custom Commands — you can wrap subagent patterns into reusable commands