GenAI Multi-Agent System

Unified architecture for multi-expert AI agent collaboration

Architecture Overview

┌─────────────────────────────────────────────────────────────┐
│                     Orchestrator                            │
│   (Manager · Lifecycle · Heartbeat ·跨机器协调)             │
└──────────────────────┬──────────────────────────────────────┘
                       │
              ┌────────▼────────┐
              │   Event Bus     │
              │     (Bell)      │
              │ Pub/Sub · 消息协议 │
              └────────┬────────┘
                       │
     ┌─────────────────┼─────────────────┐
     │                 │                 │
┌────▼────┐      ┌────▼────┐      ┌────▼────┐
│  Gen    │      │   ML    │      │ Custom  │
│  Code   │      │Researcher│     │ Agents  │
└─────────┘      └─────────┘      └─────────┘

Core Components

Component	Repo	Responsibility
Orchestrator	orchestrator	Agent lifecycle, coordination, distributed scheduling
Event Bus	bell	Pub/Sub messaging, inter-agent communication
Gen Code	gen-code	General-purpose coding agent
ML Researcher	ml-researcher	ML research agent (experiments, training, evaluation)

Project Levels

Project
 ├── Skill         # 单个能力定义
 ├── SubAgent      # 子代理配置
 ├── MCP           # Model Context Protocol 工具
 └── Agent          # 可执行实体

Design Principles

1. Event-driven — All agents communicate via Bell event bus
2. Distributed — Orchestrator can manage agents across machines
3. Hierarchical — Manager controls agent lifecycle, agents are workers
4. Extensible — New agents plug into the bus via standard protocol

Communication Flow

User/CLI
    │
    ▼
Orchestrator ──creates/destroys──► Agent Pool
    │
    │ publishes events
    ▼
Bell (Event Bus) ◄───subscribes──── Agent (GenCode / ML-Researcher / ...)
    │
    │ emits results
    ▼
Orchestrator (routes to next agent or returns to user)

License

Apache 2.0