Multi-Agent

Overview

ArgentOS supports running multiple agents, each with its own identity, configuration, model preferences, and channel bindings. This enables use cases like having a personal assistant agent, a work-focused agent, and a customer-facing agent -- all managed by the same gateway.

Defining Multiple Agents

Agents are defined in argent.json under the agents key:

{
  "agents": {
    "defaults": {
      "model": "claude-sonnet-4-20250514"
    },
    "main": {
      "name": "Argent",
      "identity": "You are Argent, a personal AI assistant."
    },
    "work": {
      "name": "Sentinel",
      "identity": "You are Sentinel, a work-focused AI that helps manage MSP operations.",
      "model": "claude-sonnet-4-20250514"
    },
    "support": {
      "name": "Helper",
      "identity": "You are Helper, a customer support agent.",
      "model": "claude-haiku-4-20250514"
    }
  }
}

Channel Binding

Each channel can be bound to a specific agent:

{
  "channels": {
    "telegram": {
      "agent": "main"
    },
    "discord": {
      "agent": "work"
    },
    "slack": {
      "agent": "support"
    }
  }
}

This means Telegram messages go to the "main" agent, Discord goes to "work", and Slack goes to "support".

Shared vs Isolated Resources

Resource	Shared?	Notes
Gateway	Shared	All agents run in the same gateway process
Memory (MemU)	Configurable	Can share or isolate memory per agent
Tasks	Configurable	Task lists can be scoped per agent
Auth profiles	Shared	All agents use the same auth profile pool
Dashboard	Shared	Switch between agents in the UI

Use Cases

Personal + Work Separation

Run one agent for personal tasks (reminders, home automation) and another for work (client management, ticket triage). Each has its own identity and memory scope.

Customer-Facing Agents

Deploy a cost-efficient agent on Haiku for customer support channels while keeping your personal agent on Sonnet/Opus.

Specialized Agents

Create agents with specific tool sets -- one with browser access for research, one with only task management for a team channel.

Limitations

All agents share the same gateway process and port
Memory isolation between agents is not yet fully implemented
The dashboard currently focuses on one active agent at a time

On this page