The system orchestrates specialized agents—a modeler, coder, and writer—working in concert to analyze problems, write code, generate mathematical models, and produce a fully formatted paper ready for submission. It includes a code interpreter (local Jupyter or cloud-based via E2B/Daytona), a RAG knowledge base backed by ChromaDB with reranking, web search via Tavily API, human-in-the-loop approval at key decision nodes, and multi-LLM support through litellm. A four-layer fault-tolerance system handles retries, fallback, evaluator shadow mode, and feedback reruns.