Multi-Agent Orchestration with Claude Code

How I set up a hub-and-spoke multi-agent system on Claude Code — reusable skill definitions, a three-stage code-quality pipeline (implement → review → resolve), and tiered routing across cloud + local LLMs.

The problem

One agent doing everything is fine for demos. For real work, you want specialization: a coder, a reviewer, a memory keeper, and an orchestrator that decides who does what. The trick is getting them to hand off cleanly without losing context.

What you'll build

• A versioned CLAUDE.md with explicit routing rules and role declarations
• Reusable skill definitions that any agent can pull in
• A three-stage code-quality pipeline: implement → review → resolve
• Tiered model routing that uses cheap/local models for routine work and reserves the big ones for hard problems

Prerequisites

• Claude Code installed and authed
• An Anthropic API key (and optionally Ollama for local routing)
• A project you want to use it on — not a toy repo

1. The CLAUDE.md contract

Coming soon — the structure I use: role declaration, routing rules, skill registry, and guardrails.

2. Defining skills

Coming soon — what makes a skill reusable vs. prompt spaghetti; how I name and version them.

3. The three-stage code-quality pipeline

Coming soon — implement → review → resolve, and why each stage has its own prompt contract.

4. Tiered model routing

Coming soon — Haiku for triage, Sonnet for implementation, Opus for review, Ollama for offline.

Wrap-up

Coming soon — what this setup actually saved me, and where it still breaks.