Aurex: Multi-Agent Orchestration Runtime for Scalable AI Coding Agents

Aurex is a multi-agent orchestration runtime developed by GeneGulanesJr (source: GitHub, link: https://github.com/GeneGulanesJr/Aurex, released on 2026-05-28) designed to coordinate AI coding agents for complex software development tasks. It features structured task decomposition, milestone management, execution guards, and specialized agent roles (Workers, Validators, Researchers, Negotiator) to enable safe, collaborative work. Key highlights include git worktree isolation, human-in-the-loop checkpoints, and a monorepo architecture with TypeScript/Fastify/React tech stack.

With the widespread adoption of AI coding assistants, teams face the challenge of coordinating multiple AI agents instead of individual work. Single agents handle code completion but struggle with complex tasks requiring architecture design, implementation, testing, and code exploration. Aurex addresses this by providing a full-lifecycle runtime for multi-agent collaboration.

System Components:

• Orchestrator: Persistent main service for task planning, milestone decomposition, and decision-making.
• PiNyx Gateway: Unified LLM access gateway (localhost:7331) for controlled model calls.
• LaPis: Shared state database with HTTP client access only.

Agent Types:

• Workers: Short-lived agents with isolated git worktrees for code execution.
• Validators: Pair with workers to assess output via immutable contracts.
• Researchers: Read-only agents for codebase exploration.
• Negotiators: Decide task outcomes (pass/retry/escalate) based on validator results.

Task Decomposition: Missions → Milestones → Working Units (hierarchical structure for traceability).

Branch Protection: Main branch is protected; workers use isolated git worktrees. Immutable Validation Contracts: Standards fixed at task start to prevent scope drift. Human-in-the-Loop: Triggers for escalation include scope changes, cost overruns, repeated failures, or validator disagreements (via Escalation Modal in frontend).

Runtime: Node.js ≥20, pnpm ≥9, TypeScript5.7. Backend: Fastify5, WebSocket (@fastify/websocket), Pi SDK, @sinclair/typebox. Frontend: React19, Vite6, Tailwind CSS4, anime.js4. Monorepo:

• shared: Types/enums/REST/WS protocols.
• backend: Server, agents, clients, orchestrator.
• frontend: App, passive/active views, custom hooks. Testing: 193 cases in 41 files (Vitest).

Prerequisites: LaPis (shared state DB) and PiNyx (LLM gateway). Installation:

1. pnpm install
2. cp .env.example .env (set endpoints/keys)
3. pnpm run build
4. Run dev (backend/frontend) or tests. Env Variables: LAPIS_ENDPOINT, PINYX_ENDPOINT, REPO_ROOT, PORT, etc.

Use Cases: Complex refactoring, feature development pipelines, code review assistance, exploratory programming, team knowledge accumulation. Value: Transitions AI from auxiliary tools to collaborative partners while maintaining human control over key decisions.

Aurex represents an evolution from single-agent tools to multi-agent systems for AI-assisted development. Its strict architecture and execution guards balance AI autonomy with human oversight, making it a valuable reference for teams looking to upgrade AI coding assistants to collaborative partners.