Agent Team Orchestration V3 (Public)

Production-proven playbook for building multi-agent content production teams on OpenClaw. Covers architecture, setup, role design, orchestration workflow, and hard-won debugging lessons.

Architecture in 30 Seconds

Single-repo + multi-role. One writer owns the real OUTPUT/ directory. All other roles (reviewers, scorer, fixer) are lightweight shell workspaces with symlinks pointing to the writer's OUTPUT. The orchestrator (main agent) dispatches work and controls flow — sub-agents never talk to each other.

Main Agent (Orchestrator)
  ├── spawn → Writer → writes to OUTPUT/
  ├── spawn → Fact-Reviewer ──┐
  ├── spawn → Style-Reviewer ─┤ read OUTPUT/ via symlinks
  ├── spawn → Data-Reviewer ──┤
  ├── spawn → Scorer ─────────┤
  └── spawn → Fixer ──────────┘ writes versioned files to OUTPUT/