The Multi-Agent Playbook

Most multi-agent demos are toys. This is the production architecture for coordinating specialized Claude agents across a real codebase: domain-specific subagents, a root orchestrator, progressive context disclosure, and clean delegation when work crosses boundaries.

• 3 principles that separate production multi-agent systems from toy demos: specialization, objectives, context disclosure
• Hub-and-spoke architecture: root .claude/ orchestrator plus per-domain agent configs with their own claude.md context
• Reusable subagent config template you drop into .claude/agents/ for any new domain (backend, frontend, db, testing)