openclaw-agent-team-orchestration（OpenClaw Agent team orchestration 效率）

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/

Five axioms:

1. File system = state center (no API, no DB)
2. AGENTS.md = role definition (auto-loaded at spawn)
3. Orchestrator = sole dispatcher (agentToAgent OFF)
4. Score threshold = quality gate (e.g., 8.5/10)
5. TOOLS.md = mandatory guardrails (prevents model tool-call loops)

Quick Start

1. Run the setup script

# Creates workspaces, symlinks, and TOOLS.md for all roles
scripts/setup-team.sh {writer-id} {fact-reviewer-id} {style-reviewer-id} {data-reviewer-id} {scorer-id} {fixer-id}

2. Register agents in openclaw.json

Add each agent to agents.list and grant spawn permissions to the main agent. The script prints the exact JSON to add.

3. Write AGENTS.md for each role

Use the templates in references/role-templates.md. The writer gets a comprehensive domain manual (400-600 lines). Reviewers, scorer, and fixer get focused role definitions (100-200 lines each).

4. Run the pipeline

Spawn writer → wait → parallel spawn reviewers → wait → spawn scorer → read score
  → pass? done. fail? spawn fixer → loop back to reviewers.

Reference Files

The Review-Fix Loop

[1] Spawn Writer → announce → verify OUTPUT/{project}/
[2] Parallel: Reviewers (fact, style, data) → all announce
[3] Spawn Scorer → announce → parse JSON from score-report.md
[4] total_score >= 8.5 → ✅ notify human
    total_score < 8.5 AND round < 3 → Spawn Fixer → back to [2]
    total_score < 8.5 AND round >= 3 → ❌ escalate to human

See references/workflow.md for spawn command patterns and edge case handling.

Critical Lessons (Save Hours)

1. TOOLS.md prevents infinite loops: Some models (k2p5 confirmed) will try to read their own AGENTS.md even though it's auto-injected, causing infinite tool-call retries. Adding three lines to TOOLS.md eliminates this completely.

2. allowAgents is required: The main agent cannot spawn sub-agents without explicit subagents.allowAgents in openclaw.json. This is the #1 setup mistake.

3. Use absolute paths in symlinks: Relative symlinks break when the working directory changes during agent context switches.

4. Version filenames explicitly in tasks: When running round 2, tell the reviewer to read article-v2.md, not "the article". Ambiguous paths cause agents to review the old version.

See references/lessons-learned.md for full details and performance benchmarks.

Scripts