AReaL-RSI is a framework for organizing a persistent research team of agents without updating backbone LLM weights. It improves performance and scalability by 2.2-36.8% compared to native scaling and Codex-Goal.