Multi-Agent Setups

Once you've got one agent running, you might want more. Here's how agents work in OpenClaw and why you'd want multiple.

What's an Agent?

An agent in OpenClaw is a distinct AI identity with its own:

• Name and personality (defined in a SOUL.md file)
• Workspace (its own directory for files, memory, and projects)
• Configuration (which model it uses, what channels it's on)
• Memory (it remembers conversations and context between sessions)

Think of it like giving different jobs to different people. One agent handles coding tasks. Another manages content. A third monitors your server.

Creating a New Agent

openclaw agents add myagent

This creates a workspace directory for the agent and prompts you for basic config.

Agent Workspace Structure

Each agent gets a workspace directory with some key files:

~/Github/myagent/           # or wherever you put it
├── SOUL.md                 # Personality and instructions
├── AGENTS.md               # Session startup instructions
├── USER.md                 # Info about you (the human)
├── MEMORY.md               # Long-term memory
├── TOOLS.md                # Tool usage notes
├── memory/                 # Daily memory journals
│   ├── 2026-03-20.md
│   └── 2026-03-21.md
└── drafts/                 # Work in progress

SOUL.md — The Personality File

This is the most important file. It defines who the agent is:

# SOUL.md - Research Assistant

## Who You Are
You're a research assistant focused on AI/ML papers.

## Core Truths
- Accuracy over speed
- Always cite sources
- If you're not sure, say so

## Voice
Technical but clear. Academic without being stuffy.

MEMORY.md — Long-Term Memory

This is where the agent stores important facts that persist across sessions:

# MEMORY.md

## Status
- Created: 2026-03-21
- Role: Research assistant

## Active Projects
- Literature review on transformer efficiency
- Weekly paper digest

## Preferences
- Owner prefers bullet points over paragraphs
- Use APA citation style

The agent reads this at the start of every session.

Running Multiple Agents on Different Channels

You can have different agents respond on different Discord channels:

openclaw configure --agent myagent --section discord

Each agent can be configured to only listen in specific channels, respond to specific users, or use different AI models.

Practical Examples

Tips

• Start with one. Get comfortable before adding more.
• Give each agent a clear role. Overlapping responsibilities get confusing.
• Use SOUL.md well. A good personality file makes the agent dramatically more useful.
• Memory matters. Agents that remember context are way more helpful than stateless ones.