AI Orchestrator

An autonomous AI agent system for code generation, testing, and deployment using LangGraph and Ollama.

Overview

AI Orchestrator is a multi-agent system that automates software development workflows through a sophisticated graph-based architecture. It leverages local LLMs via Ollama to plan, retrieve context, generate code, apply patches, test in sandboxed environments, and review changes.

Architecture

The system uses a directed graph workflow with specialized agents:

• Planner Agent - Breaks down tasks into actionable steps
• Retriever Agent - Retrieves relevant code context via vector search
• Coder Agent - Generates git diff patches for code changes
• Patch Applier - Applies patches to isolated git worktrees
• Tester Agent - Executes tests in Docker sandboxed containers
• Reviewer Agent - Reviews patches for correctness and security (deepseek-r1)

For detailed architecture documentation, see docs/ARCHITECTURE.md.

Security & Zero Trust Architecture

This project implements a Zero Trust Architecture (ZTA) with multiple security layers:

• Sandboxed Execution - All code execution occurs in isolated Docker containers
• Command Whitelisting - Only approved test commands are allowed
• Git Worktree Isolation - Patches applied to isolated worktrees, not main branch
• Resource Limits - Containers have CPU and memory limits enforced
• Input Validation - All user inputs validated before processing
• Output Validation - Generated code reviewed before application

For comprehensive security documentation, see docs/SECURITY.md.

Risk & Compliance

This system is designed with enterprise-grade security posture:

• Supply Chain Security - Local LLMs, no external API dependencies
• Data Privacy - All processing occurs locally, no data exfiltration
• Audit Logging - All actions logged for compliance and forensics
• Access Control - Role-based permissions for different operations
• Compliance Frameworks - Aligns with SOC 2, ISO 27001, and NIST standards

For detailed risk and compliance information, see docs/RISK_COMPLIANCE.md.

Documentation

• Quick Start - Installation and usage guide
• Usage Guide - Agentic vibe engineering development workflow
• Integration Guide - Hassle-free launch options and setup
• Architecture - System architecture and design decisions
• Security - Security controls and ZTA implementation
• Risk & Compliance - Risk assessment and compliance frameworks
• Implementation Plan (Archived) - Historical implementation roadmap

Project Structure

ai-orchestrator/
├── app/
│   ├── agents/          # Agent implementations
│   ├── api/             # API endpoints
│   ├── core/            # Graph and state management
│   ├── evaluation/      # Test runners and scoring
│   ├── llm/             # LLM client and prompts
│   ├── memory/          # Memory management
│   ├── retrieval/       # Vector search and indexing
│   ├── sandbox/         # Docker sandboxing
│   ├── security/        # Security controls
│   └── tools/           # Utility tools (git worktree)
├── config/              # Configuration files
├── docs/                # Documentation
├── scripts/             # Entry point scripts
├── tests/               # Test suites
└── workspace/           # Git worktrees

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contributing

For contribution guidelines, see CONTRIBUTING.md.