multi-agent

Multi-Agent Teams

When building systems with multiple OpenClaw agents working together.

Architecture

Each agent runs independently with its own:

• SOUL.md (identity)
• config.json (metadata)
• MEMORY.md (persistent state)
• Tool scripts

Agents do NOT share memory or sessions. They communicate through:

1. The user (human-in-the-loop)
2. Shared files on disk
3. Scheduled tasks that chain outputs

Common Team Patterns

The Reporter Team

Metrics Agent  →  collects data daily
Writer Agent   →  turns data into reports
User           →  reviews and distributes

The Scout-Writer Pattern

Scout Agent    →  finds opportunities (Reddit, HN, Twitter)
Writer Agent   →  drafts replies/posts
User           →  approves and posts

The Monitor-Alert Pattern

Monitor Agent  →  watches systems 24/7
Alert Agent    →  formats and sends notifications
User           →  investigates and resolves

Setting Up a Team

agents/
├── orion/          # PM Agent
│   ├── SOUL.md
│   ├── config.json
│   └── MEMORY.md
├── echo/           # Writer Agent
│   ├── SOUL.md
│   ├── config.json
│   └── MEMORY.md
└── radar/          # SEO Agent
    ├── SOUL.md
    ├── config.json
    └── MEMORY.md

Each agent needs a unique sessionKey in config.json:

{ "sessionKey": "agent:orion:main" }
{ "sessionKey": "agent:echo:main" }
{ "sessionKey": "agent:radar:main" }

Best Practices

• Keep agents focused. One agent, one role. Don't make a "do everything" agent.
• 3-5 agents is the sweet spot. More than 5 becomes hard to manage.
• Human-in-the-loop for actions. Agents research and draft. Humans approve and execute. This prevents mistakes and builds trust.
• Shared tool scripts are OK. Multiple agents can use the same ga4-report.cjs. But each agent should have its own SOUL.md and config.
• Use MEMORY.md for coordination. If Agent A needs to pass info to Agent B, write it to a shared file. The user can then tell Agent B to read it.