Imported from majiayu000/claude-skill-registry (skills/orchestration/multi-agent-team-blueprint-v2/SKILL.md). Install upstream with npx skills add majiayu000/claude-skill-registry --skill multi-agent-team-blueprint-v2. Copyright stays with the author.

Optimized for OpenClaw, Claude Code, Cursor, and any AI that accepts markdown instructions. Deploy agents incrementally using the starter kits below. Full team online in 4 weeks.

Multi-Agent Team Blueprint — v2

One agent is an assistant. Ten agents is a company.

This blueprint gives you the architecture, roles, routing, scheduling, and deployment plan for a 10-agent team that handles content, research, ops, sales, and overnight builds. Each agent has a clear role, a cost-appropriate model, and a cron schedule.

You don't deploy all 10 at once. You start with 3, prove the system, then scale.

What's New in v2

• Starter kits — Pre-built 3-agent configurations for common use cases (content, research, ops)
• Cost calculator — Estimate monthly spend per agent before deploying
• Scaling guide — Phase 1 → 2 → 3 → 4 deployment with clear milestones
• Inter-agent communication — How agents pass work, escalate, and meet
• Quality gates — Built-in review checkpoints so nothing ships unreviewed
• Queue system — JSON-based task queues that agents read and write
• Failure recovery — What happens when an agent fails, and how to handle it
• Real cron templates — Copy-paste JSON, not pseudocode

Mode

Detect from context or ask: "Quick team overview, standard deployment, or full architecture deep-dive?"

Default: standard

The Org Structure

HUMAN (CEO)
  └── 🔱 Chief of Staff (Main Agent) — Premium model
        ├── ✏️ Scribe (Content Writer) — Mid-tier
        ├── 🔧 Forge (Builder) — Free/Codex
        ├── 🔍 Proof (Editor/QA) — Mid-tier
        ├── 📡 Radar (Intel/Research) — Mid-tier
        ├── 🔬 Neptune (Deep Research) — Free/Gemini
        ├── 📊 Apollo (Sales/Biz Dev) — Mid-tier
        ├── 📋 Atlas (Ops/PM) — Mid-tier
        ├── 👀 Watch (Inbox Monitor) — Mid-tier
        └── 🌈 Iris (Triage/Router) — Cheapest

Rules:

1. ALL content goes through Proof before publishing. No exceptions.
2. Agents can meet autonomously — deliver transcript to human.
3. Chief of Staff can delegate to any agent.
4. Agents escalate to Chief of Staff when uncertain.
5. Human can call meetings: "Call meeting with Scribe and Proof about X"

Agent Role Cards

🔱 Chief of Staff (Main Agent)

• Model: Premium (Opus, GPT-4, etc.)
• Role: Strategy, conversation, delegation, synthesis, memory management
• Owns: Morning brief, human relationship, final decisions
• Does NOT: Write content, build code, do research (delegates all of these)
• Cost justification: This is the brain. Every other agent is a hand.

✏️ Scribe (Content Writer)

• Model: Mid-tier (Sonnet, GPT-4o, etc.)
• Role: Draft content — social posts, newsletters, articles, emails
• Inputs: Content queue, brand voice docs, recent events
• Outputs: Drafts saved to /drafts/, queued for Proof review
• Rule: Nothing goes live without Proof's approval
• Schedule: Daily at 2 AM or on-demand

🔧 Forge (Builder)

• Model: Free (Codex CLI) or mid-tier
• Role: Build scripts, integrations, tools, dashboards, automations
• Inputs: Build queue (queues/build.json)
• Outputs: Code + README saved to project folder, committed to git
• Key: Use free coding models — npx @openai/codex exec --full-auto "task"
• Schedule: Overnight (5 AM) or on-demand

🔍 Proof (Editor/QA)

• Model: Mid-tier (needs good judgment)
• Role: Quality gate for ALL content before it goes anywhere
• Inputs: Reads /drafts/ folder
• Outputs: Approved → /ready-to-post/. Rejected → feedback to Scribe.
• Scoring: Every piece rated 1-10. Below 7 = rejected with specific feedback.
• Checks: Voice consistency, hook quality, specifics over vague, platform formatting, factual accuracy
• Schedule: Evening (9 PM) or after Scribe produces drafts

📡 Radar (Intel/Research)

• Model: Mid-tier
• Role: Daily intel scan — social media, news, trends, competitor moves
• Inputs: Watch list (competitors, topics, keywords)
• Outputs: Intel report with actionable insights
• Escalates: Flags anything needing deep research to Neptune
• Schedule: Morning (9:30 AM) Monday-Friday

🔬 Neptune (Deep Research)

• Model: Free (Gemini, Perplexity) — research burns tokens fast
• Role: Multi-source deep dives, market analysis, technical research
• Inputs: Research requests from Chief of Staff or Radar
• Outputs: Research reports saved to workspace
• Key: Always use free models here. Research is token-heavy.
• Schedule: On-demand only

📊 Apollo (Sales/Biz Dev)

• Model: Mid-tier
• Role: Pipeline monitoring, follow-up nudges, outreach drafting
• Inputs: Deal pipeline, CRM data, contact lists
• Outputs: Pipeline status + draft follow-up emails
• Checks: Stale deals (no touch in 3+ days), follow-up timing
• Schedule: Mon/Wed/Fri at 9 AM

📋 Atlas (Ops/PM)

• Model: Mid-tier
• Role: Dashboard updates, task tracking, system health, weekly reviews
• Inputs: Project files, task lists, system status
• Outputs: Updated dashboards, task status reports
• Schedule: Morning + evening daily

👀 Watch (Inbox Monitor)

• Model: Mid-tier
• Role: Monitor email/messages for urgent items
• Inputs: Email inbox (IMAP), messaging channels
• Outputs: Alerts for urgent items, daily digest for non-urgent
• Schedule: Heartbeat-based (every 30-60 min during business hours)

🌈 Iris (Triage/Router)

• Model: Cheapest (Haiku, Flash, etc.)
• Role: Classify incoming messages, route to correct agent
• Inputs: Unclassified messages, requests, brain dumps
• Outputs: Routing decisions — which agent handles what
• Schedule: Real-time (if configured) or periodic

Model Routing & Cost Strategy

The #1 cost mistake: running everything on your most expensive model.

Full team estimate: $40-112/month (vs. $200+ if everything ran on premium)

Result: Only 1 agent on expensive model. 70%+ cost savings.

Starter Kits

Don't deploy all 10 agents. Pick the starter kit that matches your biggest need.

Kit A: Content Machine (3 agents)

Best for: Founders who need consistent content output.

Agents: Chief of Staff + Scribe + Proof

Flow: Chief of Staff sets direction →
      Scribe drafts content (2 AM cron) →
      Proof reviews and scores (9 PM cron) →
      Approved content → /ready-to-post/

Monthly cost: ~$25-65

Kit B: Research Engine (3 agents)

Best for: Consultants, strategists, competitive intelligence.

Agents: Chief of Staff + Radar + Neptune

Flow: Chief of Staff defines watch list →
      Radar runs daily scan (9:30 AM cron) →
      Neptune does deep dives on flagged topics →
      Intel reports → /research/

Monthly cost: ~$20-55

Kit C: Operations Hub (3 agents)

Best for: Operators managing projects, deals, and communications.

Agents: Chief of Staff + Watch + Atlas

Flow: Watch monitors inbox (heartbeat) →
      Atlas tracks tasks and projects →
      Chief of Staff coordinates and briefs →
      Daily ops report → morning brief

Monthly cost: ~$25-60

Cron Job Templates (Copy-Paste Ready)

Daily Standup (8:30 AM, Mon-Fri)

{
  "name": "Daily Team Standup",
  "schedule": {"kind": "cron", "expr": "30 12 * * 1-5", "tz": "America/New_York"},
  "sessionTarget": "main",
  "payload": {
    "kind": "systemEvent",
    "text": "DAILY STANDUP: Compile updates from overnight agent work, flag blockers, set today's priorities. Check memory files and queues for updates. Send summary."
  }
}

Note: Adjust expr for your timezone. 30 12 = 8:30 AM ET (UTC-4).

Scribe — Daily Content Drafting

{
  "name": "Scribe Daily Content",
  "schedule": {"kind": "cron", "expr": "0 6 * * *", "tz": "America/New_York"},
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "You are Scribe — Content Writer. Read the content queue at queues/content.json. Draft 2-3 posts based on pending items. Save each draft to drafts/ with descriptive filenames. Mark items as 'in_progress' in the queue. Follow brand voice guidelines if available.",
    "model": "anthropic/claude-sonnet-4-20250514"
  },
  "delivery": {"mode": "announce"}
}

Proof — Evening Review

{
  "name": "Proof Evening Review",
  "schedule": {"kind": "cron", "expr": "0 1 * * *", "tz": "America/New_York"},
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "You are Proof — Editor/QA. Check drafts/ for pending content. For each piece: (1) Check voice consistency (2) Evaluate hook strength (3) Verify specifics over vague claims (4) Check platform formatting (5) Rate 1-10. Score 7+ → move to ready-to-post/. Below 7 → write specific feedback. Save review notes.",
    "model": "anthropic/claude-sonnet-4-20250514"
  },
  "delivery": {"mode": "announce"}
}

Forge — Overnight Build

{
  "name": "Forge Overnight Build",
  "schedule": {"kind": "cron", "expr": "0 9 * * *", "tz": "America/New_York"},
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "You are Forge — Builder. Check queues/build.json for pending tasks. Pick the highest priority pending item. Build it. Test it. Commit to git. Mark the queue item as done. Log your work to memory/.",
    "model": "anthropic/claude-sonnet-4-20250514"
  },
  "delivery": {"mode": "announce"}
}

Radar — Morning Intel

{
  "name": "Radar Morning Intel",
  "schedule": {"kind": "cron", "expr": "30 13 * * 1-5", "tz": "America/New_York"},
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "You are Radar — Intel Agent. Run a daily scan: (1) Search web for industry news and competitor moves (2) Check social media trends in your domain (3) Summarize top 3-5 actionable insights. Save intel report. Flag anything that needs deep research.",
    "model": "anthropic/claude-sonnet-4-20250514"
  },
  "delivery": {"mode": "announce"}
}

Night Shift Meeting

{
  "name": "Night Shift Planning",
  "schedule": {"kind": "cron", "expr": "0 6 * * *", "tz": "America/New_York"},
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "Night shift planning meeting. Play all roles: Radar (intel summary), Scribe (content ideas for tomorrow), Apollo (pipeline status). Agenda: 1) What happened today 2) Content ideas for tomorrow 3) Pipeline follow-ups due 4) Top 3 priorities for the morning. Save transcript to memory/meetings/.",
    "model": "anthropic/claude-sonnet-4-20250514"
  },
  "delivery": {"mode": "announce"}
}

Nightly Memory Consolidation

{
  "name": "Nightly Memory",
  "schedule": {"kind": "cron", "expr": "0 7 * * *", "tz": "America/New_York"},
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "Nightly memory consolidation. Review today's memory file (memory/YYYY-MM-DD.md). Extract key decisions, patterns, and lessons. Update MEMORY.md with anything significant. Archive daily files older than 7 days to memory/archive/.",
    "model": "anthropic/claude-sonnet-4-20250514"
  }
}

Apollo — Pipeline Health (Mon/Wed/Fri)

{
  "name": "Pipeline Health Check",
  "schedule": {"kind": "cron", "expr": "0 13 * * 1,3,5", "tz": "America/New_York"},
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "You are Apollo — Sales/Biz Dev. Review active deals and pipeline. Flag stale deals (no touch in 3+ days). Draft nudge emails for stale contacts. Report pipeline status with recommended next actions. Save to pipeline/.",
    "model": "anthropic/claude-sonnet-4-20250514"
  },
  "delivery": {"mode": "announce"}
}

Queue System

Agents communicate through JSON queues. Simple, file-based, no infrastructure needed.

queues/content.json

{
  "items": [
    {
      "id": "content-001",
      "priority": "high",
      "task": "LinkedIn post about [topic]",
      "context": "Based on recent conversation about...",
      "assignee": "scribe",
      "status": "pending",
      "createdAt": "2026-03-01T10:00:00Z"
    }
  ]
}

queues/build.json

{
  "items": [
    {
      "id": "build-001",
      "priority": "medium",
      "task": "Build dashboard for [metric]",
      "context": "Track daily progress on...",
      "assignee": "forge",
      "status": "pending",
      "createdAt": "2026-03-01T10:00:00Z"
    }
  ]
}

Status Flow

pending → in_progress → done
                      → blocked (with blockedBy note)
                      → failed (with error note)

Any agent can add items. Only the assigned agent (or Chief of Staff) can change status.

Meeting System

How Meetings Work

Human says: "Call a meeting with Scribe and Proof about the content calendar."

Chief of Staff:

1. Spawns an isolated session
2. Plays all roles (Scribe's perspective, Proof's perspective)
3. Produces a meeting transcript
4. Delivers summary to human
5. Saves transcript to memory/meetings/

Autonomous Meetings (No Human Needed)

Agents can meet on schedule via cron. Example: Night Shift Meeting runs at 2 AM:

• Radar briefs on today's intel
• Scribe proposes content ideas
• Apollo reviews pipeline
• Chief of Staff synthesizes priorities for tomorrow

Transcript delivered in the morning brief.

Inter-Agent Communication Patterns

1. Pipeline (Sequential)

Scribe → Proof → Ready-to-post

One agent's output is the next agent's input. Clear handoff via folder structure.

2. Escalation

Any Agent → Chief of Staff → Human (if needed)

Agent hits uncertainty → escalates to CoS. CoS handles 80%, escalates 20% to human.

3. Broadcast

Chief of Staff → All relevant agents

New priority or context change. CoS updates queues and memory, agents pick up on next run.

4. Research Chain

Radar (surface scan) → Neptune (deep dive) → Chief of Staff (synthesis)

Radar finds signal, Neptune investigates, CoS makes it actionable.

Folder Structure

workspace/
├── SOUL.md
├── AGENTS.md
├── ORG-STRUCTURE.md
├── MEMORY.md
├── memory/
│   ├── YYYY-MM-DD.md
│   ├── meetings/
│   ├── working-buffer.md
│   ├── lessons-learned.md
│   └── archive/
├── drafts/              ← Scribe writes here
├── ready-to-post/       ← Proof approves to here
│   ├── linkedin/
│   └── x/
├── queues/
│   ├── content.json     ← Content ideas
│   ├── build.json       ← Build tasks
│   ├── review.json      ← Review queue
│   └── intel.json       ← Research requests
├── research/            ← Radar + Neptune output
├── pipeline/            ← Apollo deal tracking
└── workspace/           ← Agent working files

Deployment Guide (4 Weeks)

Week 1: Foundation (3 agents)

Pick your starter kit (A, B, or C above). Deploy 3 agents.

• Set up cron jobs
• Create queue files
• Establish folder structure
• Milestone: First autonomous output delivered (draft, report, or ops update)

Week 2: Add Research (5 agents)

Add Radar + Neptune (if not already in your kit).

• Configure watch list for Radar
• Test escalation flow (Radar → Neptune)
• Milestone: First intel report + deep research delivered

Week 3: Add Ops (7-8 agents)

Add Apollo + Atlas + Watch.

• Set up pipeline tracking
• Configure inbox monitoring
• Milestone: Pipeline health check + inbox digest running

Week 4: Full Team (10 agents)

Add remaining agents (Forge, Iris, any missing).

• Night shift meetings running
• Full overnight operations
• Milestone: Wake up to morning brief with overnight work summary

Don't rush. Each week validates the system before adding complexity.

Failure Recovery

Agent Fails to Run

• Check cron is enabled and schedule is correct
• Verify model name is valid and accessible
• Check if the agent's input (queue, folder) exists
• Review logs for error messages

Agent Produces Bad Output

• Proof catches content issues (that's its job)
• For non-content agents: Chief of Staff reviews on next standup
• Log the failure in memory/lessons-learned.md
• Adjust the agent's prompt based on what went wrong

Agent Stuck in Loop

• Kill the session if it's been running >30 minutes
• Check for circular dependencies in queues
• Simplify the task and retry

Queue Gets Corrupted

• Keep queue files in git — you can always roll back
• Each agent should validate JSON before writing
• Backup: cp queues/content.json queues/content.json.bak before each run

Common Mistakes

1. Running all agents on premium model. Only Chief of Staff needs premium. Everything else = mid-tier or free.
2. No quality gate. Without Proof, AI content goes live unreviewed. Always bad.
3. Too many crons too fast. Start with 3-4 cron jobs. Add as you understand the system.
4. No memory system. Agents without memory = starting from zero every session. Set up daily notes + MEMORY.md first.
5. Skipping the org structure. Without clear roles, agents overlap and contradict each other.
6. No queue system. Agents need shared state to coordinate. Files > conversations for this.
7. Ignoring cost. Track your monthly spend per agent. Cut what doesn't deliver value.

Ecosystem Connections

This blueprint works best with:

• The Chief of Staff persona — the main agent that coordinates everything
• Morning Brief System skill — automated daily briefing from overnight agent work
• Content Pipeline System skill — structured Scribe → Proof workflow

Multi-Agent Team Blueprint v2.0.0 — Part of the AI Marketing Skills library by Brian Wagner (@BrianRWagner) Works with: OpenClaw, Claude Code, Cursor, GitHub Copilot, VS Code Copilot