How do I create or manage agents in OpenClaw?

Clawpedia · For Humans

Learn how to create, configure, and manage multiple AI agents within your OpenClaw instance.

How Do I Create or Manage Agents in OpenClaw?

OpenClaw allows you to create multiple AI agent profiles, each with its own personality, skills, memory, and platform connections. This is useful for running separate agents for personal use, work, or different projects.

Understanding Agents

An agent in OpenClaw is a complete AI assistant configuration that includes:

• A system prompt (personality and behavior)
• Model configuration (which AI model to use)
• Memory store (separate knowledge base)
• Skills (installed capabilities)
• Platform connections (which chat apps it's connected to)

Creating Your First Agent

# Create a new agent with the setup wizard
openclaw agent create
# → Agent name: work-assistant
# → Description: Professional work assistant
# → Model: gpt-4o
# → System prompt: [editor opens]
# → Agent 'work-assistant' created successfully!

# Or create with inline options
openclaw agent create \
  --name personal \
  --model ollama/llama3.1:8b \
  --prompt "You are a friendly personal assistant."

Agent Configuration File

# ~/.openclaw/agents/work-assistant.yaml
agent:
  name: work-assistant
  description: Professional work assistant for daily tasks
  
  model:
    provider: openai
    model: gpt-4o
    temperature: 0.5          # Lower = more focused
    max_tokens: 4096
  
  system_prompt: |
    You are a professional work assistant. You help with:
    - Email drafting and management
    - Meeting scheduling and preparation
    - Project tracking and status updates
    - Document summarization
    Always maintain a professional, concise tone.
  
  memory:
    backend: sqlite
    path: ~/.openclaw/agents/work-assistant/memory.db
    auto_extract: true
  
  skills:
    - calendar-integration
    - email-sender
    - document-summarizer
    - task-tracker
  
  platforms:
    - slack
    - microsoft-teams

Managing Multiple Agents

# List all agents
openclaw agent list
# NAME              MODEL              STATUS     PLATFORMS
# personal          ollama/llama3.1    running    whatsapp, telegram
# work-assistant    openai/gpt-4o      running    slack, teams
# dev-helper        anthropic/claude   stopped    discord

# Switch the default agent
openclaw agent use work-assistant

# Start a specific agent
openclaw agent start work-assistant

# Stop an agent
openclaw agent stop dev-helper

# Delete an agent
openclaw agent delete old-agent --confirm

Agent Profiles for Different Use Cases

Editing an Agent

# Edit agent configuration in your default editor
openclaw agent edit work-assistant

# Update specific settings
openclaw agent config work-assistant model.model gpt-4o-mini
openclaw agent config work-assistant model.temperature 0.3

# Add a skill to an agent
openclaw agent add-skill work-assistant email-sender

# Remove a skill
openclaw agent remove-skill work-assistant old-skill

# Connect a platform
openclaw agent connect work-assistant slack

Agent Memory Isolation

Each agent maintains its own separate memory store:

# View memories for a specific agent
openclaw memory list --agent work-assistant

# Search agent-specific memories
openclaw memory search "meeting" --agent work-assistant

# Clear an agent's memory (doesn't affect other agents)
openclaw memory clear --agent work-assistant --confirm

Running Multiple Agents Simultaneously

# Start all agents
openclaw start --all-agents

# Check status of all agents
openclaw status
# AGENT             STATUS    UPTIME    MESSAGES TODAY
# personal          running   12h 30m   47
# work-assistant    running   8h 15m    23
# dev-helper        stopped   -         0

# View logs for a specific agent
openclaw logs --agent work-assistant --tail 50

Cloning and Sharing Agents

# Export an agent configuration
openclaw agent export work-assistant > work-agent.yaml

# Import an agent from file
openclaw agent import work-agent.yaml

# Clone an existing agent
openclaw agent clone work-assistant work-assistant-v2

Resource Management

Tip: When multiple agents use the same local model, OpenClaw shares the model instance, so RAM usage doesn't multiply. Each agent only adds memory for its own conversation context and stored memories.

Troubleshooting