AI Agent Orchestrator

A demo agent orchestration system showcasing tool calling, sandboxed execution, and real-time monitoring. Built with SvelteKit and TypeScript.

What This Demonstrates

This is a technical demonstration of AI agent infrastructure, not just a chat UI, but the actual orchestration layer that manages autonomous tool execution.

Key Capabilities:

• Agent execution engine with LLM-powered decision making
• Tool registry with filesystem, database, and API integrations
• Sandboxed execution with safety guardrails
• Real-time task monitoring with full audit trails
• Error handling, retries, and validation

Quick Start

npm install

Add your Groq API key to .env:

GROQ_API_KEY=your_key_here

Get a free key at console.groq.com

npm run dev

Usage Examples

Try these prompts to see the agent in action:

Simple Queries:

Query the database for all users in Engineering department

Multi-Step Tasks:

Find all engineers in the database and save their names to engineers.txt

External Data:

Get the weather in Tokyo and save it to weather.txt

File Operations:

List all files in the sandbox

Architecture

User Request → Agent Orchestrator → Tool Registry → Tool Execution → Results

• Agent Orchestrator: Manages the LLM decision loop, coordinates tools, handles errors
• Tool Registry: Validates parameters, enforces timeouts, routes to correct tools
• Tools: Sandboxed file ops, database queries (SELECT only), external APIs

Safety Guardrails

• 30s timeout per tool
• File operations limited to sandbox directory
• Database restricted to SELECT queries
• Full audit trail of all decisions

Tech Stack

• Frontend: SvelteKit, TypeScript, TailwindCSS
• Backend: SvelteKit API routes, Node.js
• LLM Provider: Groq
• Database: SQLite (better-sqlite3)
• File System: Node.js fs/promises

Known Limitations

LLM Behavior:

• Free-tier Groq API can be occasionally flaky with tool calling
• Agent may sometimes attempt extra "helpful" actions not explicitly requested
• For production use, consider Claude API or GPT-4 for better instruction adherence

Current Scope:

• Single-user demo (no authentication)
• In-memory task storage (resets on restart)
• Limited to pre-defined tools
• No multi-agent coordination

Security Note:

This is a demo with intentionally simplified security:

• Database queries use keyword blocking (production would use parameterized queries)
• No URL validation for external APIs
• No rate limiting
• No authentication

For production: Add parameterized queries, URL whitelisting, rate limiting, and authentication.