Part 1 covered the foundation: CLAUDE.md, Skills, Hooks, GitHub Actions, and Subagents. Those 5 make Claude Code smart. These next 5 make it scale.
Part 1 was “how to make one Claude session effective.” Part 2 is “how to run multiple Claudes in parallel.” Large refactors, external integrations, automated quality gates, and experimental multi-agent collaboration.
Ralph Loop
The game changer for large-scale tasks. Give Claude a job and tell it “keep going until it’s done.”
Mechanism:
1) Claude works
2) Claude tries to stop
3) Stop hook intercepts and re-injects the same prompt
4) Claude reads modified files/git history and continues
5) Loop ends on <promise>DONE</promise>
/ralph-loop "Add error handling to all API endpoints. Output <promise>DONE</promise> when complete" --max-iterations 20
Use --max-iterations as a hard safety guard.
For large migrations, pair with PROGRESS.md checklist so each loop can resume deterministically.
MCP
Model Context Protocol is the standard bridge from Claude Code to external systems.
claude mcp add github -s user -- npx @modelcontextprotocol/server-github
One command and Claude can query GitHub APIs.
High-impact use case: ticket orchestration.
Read JIRA-1234 and implement the requirements.
When done, change the ticket status to In Review and leave a comment.
Scope options:
-
-s user: global across projects -
-s local: local only, not committed -
-s project: project-level, committed/shared
Watch-outs:
- each MCP server consumes context
- MCP servers expose their own tools; they don’t inherit Read/Write/Bash
Git Worktrees
Worktrees are the infra for true parallel coding.
claude --worktree feature-social-login
claude --worktree bugfix-login-crash
Each session gets an isolated working directory + branch, so file conflicts are minimized.
Add --tmux to run in background sessions for overnight parallel batches.
claude --worktree feature-auth --tmux
claude --worktree feature-dashboard --tmux
claude --worktree refactor-api --tmux
Subagent integration:
---
name: parallel-builder
description: "Parallel build agent"
isolation: worktree
---
Code Review Plugins
Official multi-agent PR review plugins.
-
/code-review: general parallel review -
/pr-review-toolkit:review-pr: specialized review lanes
/code-review
/code-review --comment
/pr-review-toolkit:review-pr tests
/pr-review-toolkit:review-pr errors
/pr-review-toolkit:review-pr types
/pr-review-toolkit:review-pr all
Use local /code-review as pre-flight, then CI review as second pass for a strong double safety net.
Agent Teams
Experimental multi-session teamwork mode.
Subagents are one-way reporting. Agent Teams are two-way coordination between teammates.
Enable:
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
Usage:
Create 3 teammates to refactor the payment module.
API lead, DB lead, Test lead.
Share schema changes and API interface updates as you go.
Current caveats (experimental):
- unstable session resumption
- occasional premature “done” by team lead
- high token burn (multiple active model instances)
Use Agent Teams when cross-layer coordination is truly needed.
For most daily workflows, Subagents + Worktrees are enough.
Hierarchy recap:
- CLAUDE.md / Skills / Hooks = foundation
- GitHub Actions / Subagents = execution layer
- Ralph Loop / Worktrees = scaling layer
- Code Review / MCP / Agent Teams = situational tools
You don’t need all 10 at once. Start with one strong CLAUDE.md, then expand incrementally.
“Having 10 tools doesn’t mean using 10 tools. But you’d better know what all 10 do.”
Top comments (0)