AgentWrite

🤖 AgentWrite — Multi-Agent AI Writing Workbench

English | 中文

Orchestrate multiple AI agents to draft, expand, polish, and review technical articles — with personalized memory, streaming output, and async task execution.

✨ Highlights

• 🧩 Config-Driven Agent Assembly — Agents, models, prompts, MCP tools, and workflows defined in YAML, assembled at startup via Google ADK
• 📡 Streaming Output — Real-time content delivery through SSE + Redis Stream with stage-aware event processing
• 🧠 User Memory System — Mem0-inspired hybrid retrieval (Vector + BM25 + Reranker) to personalize writing with user's tech background
• ⚡ Async Long-Task Execution — Transactional Outbox + RocketMQ decouples HTTP requests from multi-agent workflows
• 🎨 Auto Diagram Generation — Dedicated Draw.io Agent creates architecture/flow/sequence diagrams inline
• 📝 Markdown Governance — CommonMark AST rendering with rule-based normalization for code fences, lists, and headings

📸 Screenshots

Login	Chat Writing	Quick Actions

Homepage	Memory Visualization

🏗️ Architecture

flowchart LR
    U[React Frontend] --> API[HTTP / SSE Controller]
    API --> APP[Application Service]
    APP --> DB[(MySQL)]
    APP --> OUTBOX[(Outbox Event)]
    OUTBOX --> RELAY[Scheduled Publisher]
    RELAY --> MQ[RocketMQ]
    MQ --> CONSUMER[Task Consumer]
    CONSUMER --> RUNNER[AgentWritingRunner]
    RUNNER --> ADK[Google ADK]
    RUNNER --> STREAM[(Redis Stream)]
    STREAM --> API
    RUNNER --> MEMORY[Memory Manager]
    MEMORY --> QDRANT[(Qdrant)]
    MEMORY --> DB
    APP --> REDIS[(Redis)]

🚀 Quick Start

Prerequisites

• JDK 17+, Maven 3.8+, Docker 20+, Docker Compose v2+
• LLM API Key (for writing model)
• Embedding API Key (for memory system)

1. Start Infrastructure

cd docs/dev-ops
docker compose -f docker-compose-environment.yml up -d

This starts MySQL (13306), Redis (16379), RocketMQ (9876/10911), and Qdrant (6333).

2. Configure API Keys

export MEMORY_EMBEDDING_API_KEY="your_embedding_api_key"
export SUTONE_WRITING_MODEL_API_KEY="your_writing_model_api_key"
export SUTONE_DRAWIO_MODEL_API_KEY="your_drawio_model_api_key"  # optional

3. Build & Run

mvn -pl sutone-agent-bok-app -am clean package -Dmaven.test.skip=true
java -jar sutone-agent-bok-app/target/sutone-agent-bok-app.jar

The app runs at http://localhost:8091.

📦 Module Structure

sutone-agent-bok
├── sutone-agent-bok-api              # External API interfaces, DTOs, response objects
├── sutone-agent-bok-types            # Shared enums, exceptions, base types
├── sutone-agent-bok-domain           # Domain entities, services, repositories, Agent logic
├── sutone-agent-bok-infrastructure   # MySQL, Redis, Qdrant, external service adapters
├── sutone-agent-bok-trigger          # HTTP, SSE, RocketMQ Consumer, scheduled tasks
└── sutone-agent-bok-app              # Spring Boot entry point & configuration assembly

The backend follows a DDD-style layered architecture, isolating domain logic from infrastructure adapters (MySQL, Redis, Qdrant, RocketMQ).

🛠️ Tech Stack

Area	Technology
Backend	Java 17, Spring Boot 3.4.3, Maven
Agent	Google ADK 0.5.0, Spring AI, MCP, RxJava
Storage	MySQL, Redis / Redisson, Qdrant
Async & Streaming	RocketMQ, Redis Stream, SSE
Data Access	MyBatis, HikariCP
Deployment	Docker, Docker Compose

🗺️ Roadmap

• Automated evaluation suite and quality metrics for Agent workflows
• Task trace, model latency, and token consumption monitoring
• Outbox retry, task recovery, and consumer idempotency tests
• Deployment documentation and API examples

🤝 Contributing

Contributions are welcome! Please read CONTRIBUTING.md for guidelines.

📄 License

This project is licensed under the MIT License.

⭐ Star History