I open-sourced a multi-agent writing system. You give it a draft, and 6 agents work in sequence: analyze requirements, generate content, review quality, identify where diagrams belong, generate Draw.io XML, then inject diagrams back into the article.
What makes it different from a ChatGPT wrapper:
- Agents are config-driven. Each agent's model, prompt, tools, and workflow are defined in YAML files, assembled at startup. Swapping an agent or adding a new one is a config change.
- It has memory. A Mem0-inspired system stores user tech background and writing preferences in Qdrant, retrieves them via hybrid search (vector + BM25 + Reranker), and injects relevant context into prompts.
- Long tasks run async. The system uses Transactional Outbox + RocketMQ so agent workflows don't block HTTP requests. Results stream back via SSE.
Built with Java 17, Spring Boot 3.4.3, Google ADK 0.5.0. Full source with Docker Compose setup:
[link] [comments]