Multi-Agent AI Orchestration

💡 Real story: In 14 days, Fu Sheng's single AI lobster "Sanwan" evolved into 8 specialized agents. The Commander dispatches tasks via Feishu; the Writer produces content; the Strategist does research; the Ops agent handles daily routines; the Evolution agent writes code; the Community agent manages Discord; the Trading agent monitors stocks. Together they run 7×24 — the CEO just says one sentence in the morning.

🏛️ Architecture Overview

Multi-agent = one OpenClaw installation × multiple independent roles, each with its own persona, memory, and skill set

🎯 Commander (Main Agent)

↙ ↓ ↘

✍️ Writer 🔬 Strategist 📋 Ops Agent 💻 Dev Agent 💬 Community

Spawns sub-agents · Shares files for collaboration · Push-based result reporting

🧩 Why Multiple Agents?

🎯 Specialization beats generalization
A dedicated writer agent with a journalism-style SOUL.md produces far better content than a generalist. Each agent can be tuned for exactly one domain.

⚡ Parallel execution
While the Writer drafts an article, the Strategist is doing competitive research, and the Ops agent is checking messages. True parallelism — not sequential.

🧠 Context isolation
Long-running agents don't pollute each other's context. The trading agent's 10,000 token market analysis doesn't crowd out the writer's creative flow.

🔒 Role-based security
The community-facing agent only sees public information. The financial agent accesses sensitive APIs. Least-privilege per role.

👥 Classic Role Designs

🎯 Commander — The Orchestrator

Receives tasks from humans (via Feishu/Discord), breaks them into sub-tasks, delegates to the right specialist, consolidates results, and reports back. The only agent the human directly talks to.

✍️ Writer — Content Specialist

Long-form articles, social posts, email drafts. Trained on the owner's writing style via SOUL.md examples. Never does research itself — relies on Strategist to supply facts.

🔬 Strategist — Research & Analysis

Web searches, competitor analysis, market research, data synthesis. Produces structured reports for other agents to consume. No writing, no publishing — just intelligence.

📋 Ops Agent — Daily Execution

Calendar, emails, reminders, routine tasks. Runs heartbeat for operational workflows. The agent equivalent of an executive assistant.

💻 Evolution Agent — Code & Systems

Writes and deploys code, maintains websites, monitors infrastructure, executes Claude Code sessions. Technical depth that other agents don't need.

🔗 Communication Patterns

Pattern	How It Works	Use Case
Spawn & Report	Commander spawns a sub-agent with a task; sub-agent auto-announces result when done	One-off complex tasks
Shared Files	Agents read/write shared markdown files in workspace as a bulletin board	Async collaboration, handoffs
Feishu Relay	Agents send results to a shared Feishu channel; others subscribe and react	Human-in-the-loop approval
Heartbeat Poll	Each agent runs its own heartbeat, checking a shared task queue file	Parallel autonomous execution

🛠️ Building Your First Multi-Agent System

Map your workflows to roles

List everything your current single agent does. Group tasks by domain: content vs research vs code vs ops. Each major domain = one potential agent.

Create a workspace per agent

Each agent gets its own workspace directory with a SOUL.md, MEMORY.md, and HEARTBEAT.md. Shared resources go in a common shared/ folder.

workspaces/
  commander/   SOUL.md  MEMORY.md  HEARTBEAT.md
  writer/      SOUL.md  MEMORY.md  HEARTBEAT.md
  strategist/  SOUL.md  MEMORY.md  HEARTBEAT.md
  shared/      tasks.md  context.md  results.md

Define hand-off protocols

Specify exactly how agents pass work to each other. A simple convention: the Commander writes to shared/tasks.md; specialists poll it during heartbeat.

# shared/tasks.md
## [PENDING] Strategist: Research top 5 AI coding tools
- Output: shared/research-ai-coding.md
- Requested by: Commander

## [DONE] Writer: Draft article on Claude 3.5 Sonnet
- Output: shared/articles/claude-35.md
- Status: ✅ Published 2024-03-14

Start simple, evolve gradually

Begin with Commander + one Specialist. Prove the workflow works for 1 week before adding the next agent. Don't design 8 agents on day 1.

⚠️ Common Pitfalls

❌ Too many agents too soon — Each agent needs its own SOUL.md, memory, and maintenance. Start with 2, not 8.

❌ No clear ownership — If two agents can both "handle" a task, neither will do it reliably. One task, one agent.

❌ Over-coupling — Agents that can't work without others are fragile. Each agent should be useful standalone; collaboration is a bonus.

✅ Right pattern — Start with Commander + Writer. The Writer publishes content autonomously. The Commander handles everything else until a second specialist is needed.

💡 The 14-day rule: Fu Sheng's team grew from 1 to 8 agents in 14 days — but each agent was only added when a specific bottleneck appeared. You'll know when you need a new agent when one existing agent is constantly overloaded with tasks from one specific domain.

🔗 Related Skills

🏗️ Multi-Agent Team Setup — Step-by-step configuration guide with code examples

💓 Heartbeat Mechanism — Keep each agent running autonomously on a schedule

🧬 SOUL.md Design — Define distinct personalities and expertise for each agent role

🤝 Multi-Agent Orchestration

🏛️ Architecture Overview

🧩 Why Multiple Agents?

👥 Classic Role Designs

🔗 Communication Patterns

🛠️ Building Your First Multi-Agent System

⚠️ Common Pitfalls

🔗 Related Skills