The paper presents Auto-RecSys, an autonomous research-agent system for running large-scale recommender-system experimentation, using distributed asynchronous execution to run multiple experiments in parallel across servers with centralized memory for persistent state across failures. It implements a dual-loop architecture: an Execution Evolution Loop that accumulates operational knowledge from recorded failures and successful pipelines, and an Idea Evolution Loop that uses experiment results to steer future research directions. The design separates natural-language guidance for LLM reasoning from deterministic scripts that enforce operational reliability in resource-intensive production settings.
