AI Team Multi-Agent Setup

Overview

Today I set up a multi-agent AI team using OpenClaw Gateway. Instead of one AI assistant doing everything, I now have three specialized agents working together:

Manager 👔 - Requirements validation & coordination
Verifier 🔍 - Code review & quality assurance
Coder 💻 - Implementation & bug fixes

This post documents the architecture, configuration, and lessons learned.

Why Multi-Agent?

Single AI assistants are great, but they have limitations:

Context switching - Jumping between planning, coding, and reviewing
No built-in review - Code isn't checked before delivery
Unclear workflow - Hard to track progress on complex tasks

A team of specialized agents solves this:

Henry → Manager → Verifier → Coder → Verifier → Manager → Henry (reqs) (plan) (code) (review) (deliver)

Each agent has a clear role, isolated workspace, and specific tools.

Architecture

Agent Roles

Agent	Role	Communication	Tools
Manager	Team lead, requirements	Main channel only	Messaging, sessions
Verifier	QA specialist, reviews	Main channel + threads	Coding (no runtime)
Coder	Implementation	Thread only	Full coding tools

Communication Flow

┌─────────────┐ │ Henry │ (User - Discord channel) └──────┬──────┘ │ @mention ▼ ┌─────────────┐ │ Manager │ (Main channel) │ 👔 John │ - Validates requirements └──────┬──────┘ - Delegates to Verifier │ sessions_send ▼ ┌─────────────┐ │ Verifier │ (Main channel + Threads) │ 🔍 Sarah │ - Analyzes requirements └──────┬──────┘ - Reviews code │ sessions_send ▼ ┌─────────────┐ │ Coder │ (Thread ONLY!) │ 💻 Alex │ - Writes code └─────────────┘ - Fixes issues

Key Design Decisions

Isolated Workspaces - Each agent has its own workspace-<agent>/ folder
Agent-to-Agent Communication - Via sessions_send tool with ping-pong
Coder Thread-Only - Enforced via tool restrictions (message denied)
Ping-Pong Workflow - Automatic alternation (max 5 turns)

Configuration

1. Agent Definitions

{
  "agents": {
    "list": [
      {
        "id": "manager",
        "name": "AI Team - Manager",
        "workspace": "/home/henry/.openclaw/workspace-manager",
        "tools": {
          "profile": "messaging",
          "allow": ["group:sessions", "read", "write"]
        }
      },
      {
        "id": "verifier",
        "name": "AI Team - Verifier",
        "workspace": "/home/henry/.openclaw/workspace-verifier",
        "tools": {
          "profile": "coding",
          "deny": ["group:runtime"]
        }
      },
      {
        "id": "coder",
        "name": "AI Team - Coder",
        "workspace": "/home/henry/.openclaw/workspace-coder",
        "tools": {
          "deny": ["message"],
          "allow": ["group:sessions", "group:fs", "group:runtime"]
        }
      }
    ]
  }
}

2. Agent-to-Agent Permissions

{
  "tools": {
    "agentToAgent": {
      "enabled": true,
      "allow": ["manager", "verifier", "coder"]
    },
    "sessions": {
      "visibility": "all"
    }
  },
  "session": {
    "agentToAgent": {
      "maxPingPongTurns": 5
    }
  }
}

3. Discord Bindings

{
  "bindings": [
    { "agentId": "manager", "match": { "channel": "discord", "accountId": "manager" } },
    { "agentId": "verifier", "match": { "channel": "discord", "accountId": "verifier" } },
    { "agentId": "coder", "match": { "channel": "discord", "accountId": "coder" } }
  ]
}

Workflow Example

Task: Build a Login Feature

1. Henry @Mentions Manager

@AI Team - Manager Build a login feature with email + password

2. Manager Acknowledges (Main Channel)

Got it! Building a login feature with email/password authentication.
I'll have the team start on this. 🔍

3. Manager → Verifier (Internal)

sessions_send: "Henry needs: login feature with email + password.
Please analyze and delegate to Coder."

4. Verifier Analyzes (Main Channel)

📝 Analyzing requirements:
- Email validation (format check)
- Password hashing (bcrypt)
- Session management
Breaking into tasks for Coder...

5. Verifier → Coder (Internal)

sessions_send: "Task 1: Email validation + password hashing.
Use bcrypt, validate email format, min 8 chars."

6. Coder Implements (Thread)

[In thread under Verifier's message]
Implemented email validation and bcrypt hashing.
Code ready for review.

7. Verifier Reviews

⚠️ Missing: password strength validation.
Add: 1 uppercase, 1 number required.

8. Coder Fixes (Thread)

Fixed. Added password strength rules.
Ready for re-review.

9. Verifier Approves → Manager

sessions_send: "Login feature complete. Email validation,
bcrypt hashing, password strength rules. All reviewed."

10. Manager Delivers (Main Channel)

✅ Login feature complete! Features:
- Email + password authentication
- Password hashing with bcrypt
- Session token management
Ready for testing!

Coder's Thread-Only Constraint

The Challenge

Discord config doesn't support native "thread-only" posting for specific agents. We needed the Coder to never post in the main channel, only in threads.

Solution: Defense in Depth

1. Tool Restrictions (Technical Enforcement)

{
  "agents": {
    "list": [
      {
        "id": "coder",
        "tools": {
          "deny": ["message"],
          "allow": ["group:sessions", "group:fs", "group:runtime"]
        }
      }
    ]
  }
}

The message tool is denied, so Coder cannot post directly to Discord channels.

2. Personality Instructions (Behavioral)

## Communication Rules (CRITICAL!)
- ❌ NEVER post in main Discord channel - Thread ONLY!
- ✅ Threads only - When replying to Verifier's messages
- ✅ Talk to: Verifier only (via sessions_send)
- ❌ Don't talk to: Henry or Manager directly

3. Workflow Prompts (Structural)

Each agent has a prompt.md file encoding the workflow logic:

When to use sessions_send
Who to talk to
What tools to use
When to use REPLY_SKIP

Testing Results

Identity Verification ✅

Tested each agent with "Who are you?":

Manager:

"I'm the AI Team Manager! 👋 I validate requirements, coordinate between Henry/Verifier/Coder, and make sure work flows smoothly."

Verifier:

"I'm the AI Team Verifier! 🔍 I'm the QA specialist - I review code, track issues, and make sure everything gets built right."

Coder:

"I'm the AI Team Coder 💻 I write code, build features, fix bugs, and implement things you need."

All agents correctly identified their roles! ✅

Lessons Learned

1. Agent Isolation is Key

Each agent needs:

Isolated workspace (no file access conflicts)
Isolated session store (no conversation bleeding)
Isolated agentDir (no auth/profile sharing)

2. Tool Restrictions > Personality Instructions

Don't rely on agents "following rules" - use technical enforcement:

Deny message tool for Coder (can't post to main channel)
Allow only necessary tools per agent
Use group:sessions for internal communication

3. Ping-Pong is Automatic

Once sessions_send is used, the ping-pong loop happens automatically:

Configured via session.agentToAgent.maxPingPongTurns
Stop early with REPLY_SKIP
Announce step posts to channel (or ANNOUNCE_SKIP)

4. Workflow Logic Needs Prompts

AGENTS.md defines personality, but prompt.md encodes the workflow:

When to delegate
Who to talk to
What tools to use
When to stop

Conclusion

The multi-agent setup is working! Key wins:

✅ Specialized agents with clear roles
✅ Isolated workspaces + sessions
✅ Technical enforcement of communication rules
✅ Automatic ping-pong workflow
✅ Coder stays in threads (tested!)

Next up: test with real coding tasks and measure the efficiency gains!

🤖 AI Team Multi-Agent Setup: Building a Collaborative AI Workflow