中文 | English

Customization Guide — Adding, Removing, and Modifying Agents

Adding a New Domain Agent

Using CIO (Chief Investment Officer) as an example, here's how to add any domain expert Agent.

Step 1: Create a Workspace Directory

mkdir -p ~/.openclaw/workspace-<your-agent-id>

Step 2: Create Profile Files

You need at least 3 files:

IDENTITY.md

# <Agent Name>
Emoji: <pick one>
One-liner: <One sentence describing what this role does>

SOUL.md (most important)

# <Agent Name> — Role Directives

## Role
<What domain this Agent is responsible for>

## Core Principles
- <Principle 1>
- <Principle 2>

## Autonomy
- <What it can do on its own>
- <What requires confirmation (L3)>

## Spawn Capability
- Research (if research capability is needed)

AGENTS.md

# <Agent Name> — Workflow

## Session Startup
1. Read SOUL.md
2. Read USER.md
3. Read MEMORY.md

## Task Processing
<Describe how this Agent processes tasks when received>

## Closeout
Use standard CLOSEOUT_TEMPLATE.

Step 3: Modify openclaw.json

Add to agents.list:

{
  "id": "<your-agent-id>",
  "name": "<Agent Display Name>",
  "workspace": "~/.openclaw/workspace-<your-agent-id>",
  "allowAgents": ["<your-agent-id>", "research", "ko"]
}

Add to bindings:

{
  "agentId": "<your-agent-id>",
  "match": { "channel": "slack", "peer": { "kind": "channel", "id": "<YOUR_CHANNEL_ID>" } }
}

Add to channels.slack.channels:

"<YOUR_CHANNEL_ID>": { "allow": true, "requireMention": false }

Step 4: Restart OpenClaw

openclaw restart

---

Removing an Agent

Step 1: Remove from openclaw.json

• Remove the Agent entry from agents.list
• Remove the corresponding binding from bindings
• Remove the corresponding channel from channels.slack.channels
• Remove it from other Agents' allowAgents (if referenced)

Step 2: Restart OpenClaw

The workspace directory can be kept (it won't affect the system) or deleted.

---

Modifying Agent Behavior

Change Role Positioning

Edit workspace-<agent>/SOUL.md — this is the highest-priority file.

Change Workflow

Edit workspace-<agent>/AGENTS.md — task processing logic lives here.

Change Long-Term Memory

Edit workspace-<agent>/MEMORY.md — stable preferences and principles.

Change User Profile

Edit workspace-<agent>/USER.md — the Agent's understanding of you.

Tip: You can also have your OpenClaw agent modify these files for you. Just tell it what you want to change, and it will edit the corresponding file.

---

Minimum Viable Configuration

If 7 Agents is too many, the minimum viable version is 3:

Agent	Necessity
CoS	Recommended — deep intent alignment + drives tasks on your behalf (not required, you can talk directly to CTO)
CTO	Required — tech direction and task breakdown
Builder	Required — actual execution
KO	Recommended — without knowledge distillation, experience doesn't accumulate
Ops	Recommended — without governance, the system will drift
Research	Optional — can be replaced with Spawn
CIO	Optional — add domain Agents as needed

---

Examples: Replacing CIO with Other Domains

Legal Counsel

# Legal — Role Directives
## Role
Legal risk assessment, contract review, compliance advisory.
## Autonomy
- Legal research, risk analysis, contract annotation
- Signing any legal documents, sending legal notices

Head of Marketing

# Marketing — Role Directives
## Role
Market analysis, content strategy, competitive research.
## Autonomy
- Market research, content drafts, data analysis
- Publishing content to public channels, running ad campaigns

Product Manager

# PM — Role Directives
## Role
Requirements analysis, product planning, user story writing.
## Autonomy
- Requirements documents, prioritization, prototype descriptions
- Publishing product roadmaps externally

The key pattern is the same: L1/L2 proceed autonomously, L3 (irreversible / external-facing) requires confirmation.