RecSys Factory: Bounding LLM Agent Autonomy in Industrial Recommenders

A recent paper on arXiv (2608.11241) presents RecSys Factory, a platform utilizing LLM-agent technology that has been operational for 78 days across three different Tencent recommender business sectors. The authors tackle what they term the 'autonomy-determinism-efficiency trilemma' in industrial recommendation systems, focusing on general autonomy (interpreting operator intent and generating glue code without prior examples), industrial determinism (ensuring schema-compliant feature extraction and stable A/B testing, with no compliance-path hallucinations), and overall efficiency. The design limits autonomy to specific decision points, achieved through three deconstructions that address each aspect of the trilemma. The runtime is divided into three event sources: Claude Code Stop hooks, corporate-IM webhooks, and workflow scheduler APIs. Notably, the platform does not run any long-lived daemons during idle periods, utilizing zero CPU during 94% of its downtime. The paper positions this deployment as a pragmatic approach to balancing the needs of general-purpose LLM agents with the stringent demands of industrial systems, relevant to AI technology and digital infrastructure, though not focused on art.