Multi-Agent Workflows
Run multiple sub-agents in parallel instead of sequentially.
Run multiple sub-agents in parallel instead of sequentially.
What it is
Sub-agents are independent Claude instances with their own context. The main agent orchestrates, the sub-agents do the actual work in parallel. Result: split complex tasks, run them at the same time, and keep the main context clean and cheap.
Install / Setup
Sub-agents are built into Claude Code (Task tool / sub-agent tool) โ nothing to install. You just need the right instruction so Claude actually uses them.
Option 1 โ CLAUDE.md rule:
## Sub-agents (IMPORTANT โ ALWAYS use)
- Default is a sub-agent โ anything beyond a quick Grep/Edit
- Code changes โ sub-agent reads, understands, edits. Main context stays clean.
- Multi-file edits โ multiple sub-agents in parallel
- Research โ sub-agent
- Main context only for: planning, decisions, short status updates
- Parallelize โ fire independent sub-agents at the same time, not one after another
Option 2 โ Prompt Claude Code:
Add a sub-agent strategy to my ~/.claude/CLAUDE.md: default is sub-agent for anything beyond a quick Grep/Edit, multi-file edits in parallel, main context only for planning/decisions. Check first if there's already something about sub-agents and merge cleanly.
Or ad-hoc in chat:
Fire off 3 parallel sub-agents: one documents all API routes, one checks the DB schemas, one reviews the tests. Summarize the results for me.
Liking this?
Inside the community I show you how I run all of this day-to-day โ live sessions, direct feedback on your setup, and my full configs.
Use cases
- Multi-file refactor โ 5 sub-agents edit one file each in parallel instead of 10 minutes sequentially.
- Codebase recon โ A sub-agent "grep + read + understand" hands you a focused result and keeps your main context empty.
- Competitor analysis โ 3 sub-agents analyze 3 competitor sites at once.
- Test generation for a whole module โ One sub-agent per file writing tests, main agent reviews.
- Docs for a big system โ Parallel sub-agents, one per subsystem.
Pro tip
Every sub-agent burns tokens. Don't fire one off for every tiny task. Rule of thumb: if the task is a clearly scoped research/edit with a structured output โ sub-agent. If you're steering and deciding the whole time โ main context.
๐ Inside the community we build complete automation systems and workflows that actually run your day. โ Join the Claude Code Mastery Community
Updated regularly โ follow @vine.codes for more.
Want more?