AI Agent Managing Dashboard (with kill switch)

https://github.com/user-attachments/assets/9281c4d1-d2a5-49dc-addc-9733a5a503ea

A real-time dashboard for monitoring and terminating mock AI agents built with Phoenix and Elixir. I wanted to gain some familiarity with Elixir and Svelte.

Overview

This application simulates multiple AI agents that periodically emit thoughts and actions. A web dashboard allows operators to monitor agent behavior in real-time and instantly kill misbehaving agents.

Features

• Mock AI Agents: GenServer-based agents that simulate AI behavior with periodic heartbeats
• Real-time Monitoring: Phoenix Channels provide live updates of agent status and logs
• Kill Switch: HTTP API and real-time UI for terminating agents
• Process Supervision: DynamicSupervisor manages agent lifecycles
• Registry-based Discovery: Efficient agent lookup using Elixir Registry

Architecture

Backend (Elixir/Phoenix)

• AgentKillSwitch.Agents.MockAgent: GenServer simulating AI agent behavior
• AgentKillSwitch.Agents.Supervisor: DynamicSupervisor managing agent processes
• AgentKillSwitchWeb.AgentChannel: Phoenix Channel for real-time updates
• AgentKillSwitchWeb.AgentController: HTTP API for agent spawn/kill operations

Frontend (Svelte - TODO)

• Real-time dashboard showing agent status
• Agent logs and kill buttons
• WebSocket connection to Phoenix Channels

Getting Started

Prerequisites

• Elixir 1.14+
• Phoenix 1.7+

Installation

# Install dependencies
mix deps.get

# Run tests
mix test

# Start the server
mix phx.server

The application will start 3 mock agents automatically and be available at http://localhost:4000.

API Endpoints

• POST /api/agents - Spawn a new agent
• POST /api/agents/:id/kill - Kill an agent by ID

WebSocket Channel

• Topic: agents:lobby
• Messages: agent_update, agent_terminated

Configuration

Agent behavior can be configured in AgentKillSwitch.Agents.Supervisor:

• tick_interval_ms: Agent heartbeat interval (default: 1000ms)
• max_log_entries: Maximum log entries per agent (default: 20)
• auto_respawn: Whether agents respawn after being killed (default: false)

Testing

Run the test suite:

mix test

Tests cover:

• Agent lifecycle management
• Supervisor operations
• State transitions and log management

Architecture Notes

• Agents run continuously regardless of client connections
• Process termination is immediate but allows current broadcasts to complete
• Registry provides O(1) agent lookup by ID
• PubSub enables efficient broadcasting of agent updates