Multi-Agent Setup

This guide is currently available in Chinese only. Read in Chinese →

Learn how to split one OpenClaw instance into multiple specialized Agents — a writing assistant, a news aggregator, a task coordinator — each with isolated memory but shared context.

Key Concepts

Multi-instance vs Multi-Agent: Multi-instance = separate processes for fault isolation. Multi-Agent = multiple roles within one process for lightweight task division. The two complement each other.

Agent isolation: Each Agent has its own working directory, session state, and memory files. Physical-level isolation — agents can't pollute each other's memory even when sharing a Telegram bot.

Workspace strategy: Share the workspace so agents inherit your style files (SOUL.md, USER.md), but keep state directories isolated per agentId automatically.

Setup Steps

1. Create Telegram groups for each agent role
2. Add your bot to each group
3. Tell your main OpenClaw agent the group IDs and describe each role — it handles the rest

Common Pitfall: Telegram Group Privacy

By default, bots only see messages where they're @mentioned. Fix: go to @BotFather → /setprivacy → Disable. Or make the bot a group admin (admins bypass privacy mode).

See the full guide in Chinese for

• Architecture diagram (multi-instance + multi-agent layers)
• Workspace sharing vs independent decision guide
• Two interaction modes: direct group chat vs main agent delegation
• Model assignment by task type to reduce costs
• Team scaling rhythm recommendations