Multi-Agent Communication

RAPID includes a built-in event bus that allows multiple AI agents to communicate and coordinate their work on the same project.

Overview

The event bus enables scenarios like:

• Specialist collaboration: Have Claude handle architecture while Aider focuses on tests
• Parallel development: Multiple agents working on different features simultaneously
• Review workflows: One agent writes code, another reviews it
• Task delegation: Agents can request help from other agents

Quick Start

1. Enable the event bus (enabled by default):

   rapid.json

   {
     "eventBus": {
       "enabled": true
     }
   }

2. Start multiple agents:

   # Terminal 1
   rapid dev
   
   # Terminal 2
   rapid dev --agent opencode

3. Check connected agents:

   rapid bus agents

How It Works

The event bus uses Redis (managed automatically in Docker) to provide:

• Persistent messaging: Messages are stored and can be retrieved by agents that connect later
• Agent presence: Track which agents are currently active
• Message routing: Send messages to specific agents or broadcast to all

Architecture

flowchart LR
    Claude["Claude Agent"]
    OpenCode["OpenCode Agent"]
    Aider["Aider Agent"]
    Redis["Redis Event Bus"]

    Claude <--> Redis
    OpenCode <--> Redis
    Aider <--> Redis

MCP Tools for Agents

When the RAPID MCP server is connected, agents have access to these event bus tools:

bus_register

Register the agent with the event bus.

// Called automatically when agent starts
bus_register({ name: 'claude', capabilities: ['code', 'review'] });

bus_send

Send a message to other agents.

// Send to all agents
bus_send({
  type: 'task_complete',
  content: 'Finished implementing auth module',
});

// Send to specific agent
bus_send({
  type: 'review_request',
  content: 'Please review src/auth.ts',
  to: 'opencode',
});

bus_messages

Get recent messages from the bus.

// Get all messages
bus_messages({ limit: 50 });

// Get messages from specific agent
bus_messages({ from: 'aider', limit: 10 });

bus_agents

List currently connected agents.

bus_agents();
// Returns: [{ id: "claude-abc", name: "claude", lastSeen: "..." }, ...]

Message Types

Common message types for agent coordination:

Type	Description
task_start	Agent starting a task
task_complete	Agent finished a task
task_blocked	Agent blocked, needs help
review_request	Request code review
review_complete	Review finished
question	Ask other agents a question
answer	Response to a question
file_changed	Notify about file changes
test_result	Share test results

Configuration

Basic Configuration

rapid.json

{
  "eventBus": {
    "enabled": true
  }
}

Advanced Configuration

rapid.json

{
  "eventBus": {
    "enabled": true,
    "redis": {
      "port": 6379,
      "dataDir": "~/.rapid/redis-data"
    },
    "injection": {
      "enabled": true,
      "promptTemplate": "Connected agents: {{agents}}"
    }
  }
}

Configuration Options

Option	Default	Description
enabled	true	Enable/disable the event bus
redis.port	6379	Redis port
redis.dataDir	none	Persist Redis data
injection.enabled	true	Inject agent info into prompts
injection.promptTemplate	default	Custom prompt template

CLI Commands

Check Status

rapid bus status

Output:

RAPID Event Bus Status
────────────────────────────────

Project:    my-project
Mode:       Redis (persistent)
URL:        redis://localhost:6379
Container:  d5263f2e5c71

Messages:   42
Agents:     3

View Message History

# Recent messages
rapid bus history

# Last 10 messages
rapid bus history --limit 10

# Messages from specific agent
rapid bus history --agent claude-abc123

# Filter by type
rapid bus history --type review_request

List Connected Agents

rapid bus agents

Output:

Connected Agents
────────────────

  claude-abc123
    Name:         claude
    Connected:    2 hours ago
    Last active:  just now
    Capabilities: code, review, test

  opencode-def456
    Name:         opencode
    Connected:    1 hour ago
    Last active:  5 minutes ago
    Capabilities: code, refactor

Workflows

Code Review Workflow

1. Claude writes code:

   Claude: Sends message type "review_request" with file path

2. OpenCode reviews:

   OpenCode: Receives message, reviews code, sends "review_complete"

3. Claude addresses feedback:

   Claude: Receives review, makes changes

Parallel Feature Development

1. Start both agents on different features:

   # Terminal 1 - Claude on auth
   rapid dev
   # Claude works on src/auth/
   
   # Terminal 2 - Aider on API
   rapid dev --agent aider
   # Aider works on src/api/

2. Agents notify each other of changes:

   Claude: "file_changed" - src/auth/types.ts
   Aider: Sees notification, updates API imports

Task Delegation

1. Claude encounters a task better suited for another agent:

   Claude: Sends "task_blocked" with context

2. OpenCode picks up the task:

   OpenCode: Receives message, handles the task
   OpenCode: Sends "task_complete" when done

Best Practices

1. Use clear message types: Stick to standard types for interoperability
2. Include context: Messages should have enough context to act on
3. Avoid conflicts: Coordinate which agent works on which files
4. Monitor the bus: Use rapid bus status to check health
5. Clean up: Messages are cleared when you restart the event bus

Troubleshooting

Event bus not starting

# Check if Docker is running
docker ps

# Check Redis container
docker logs rapid-redis

# Manually restart
rapid stop --remove
rapid dev

Agents not seeing messages

# Verify agents are registered
rapid bus agents

# Check message history
rapid bus history --limit 100

High memory usage

Redis stores all messages in memory. If you have many messages:

# Stop and remove to clear messages
rapid stop --remove

# Or configure data persistence
# rapid.json: eventBus.redis.dataDir = "~/.rapid/redis"

Disabling the Event Bus

If you don’t need multi-agent features:

rapid.json

{
  "eventBus": {
    "enabled": false
  }
}

This skips Redis startup and event bus initialization.