RecSys Factory Balances Autonomy, Determinism, and Efficiency in Recommenders

Deploying large language model (LLM) agents into industrial recommender systems presents a complex challenge, balancing the need for general autonomy with the demands of industrial determinism and end-to-end efficiency. This tension, termed the autonomy-determinism-efficiency trilemma, suggests that maximizing any two aspects often compromises the third. Tencent has introduced RecSys Factory, an LLM-agent platform designed to navigate this trilemma. RecSys Factory's core principle is to confine agent autonomy to specific decision points within the recommender lifecycle, rather than granting control over entire pipelines. This is achieved through three key deconstructions: runtime is broken down into host-emitted events, minimizing CPU usage during wait times; capability is structured into a comprehensive skill ecosystem with pitfall tables, bounding autonomy to typed decision surfaces; and deployment is made adaptable across diverse business lines with differing semantics and A/B topologies. The platform, which has been deployed for 78 days across three Tencent business lines, recorded 1,624 CLI-tool dispatches with a 78.6% success rate. It also incorporates a human-in-the-loop card protocol for diagnostics and audit trails, ensuring human oversight at critical junctures. While observing onboarding-time compression in some cases, the findings are presented as case-study observations rather than generalized claims, highlighting a practical approach to integrating LLM agents into complex industrial systems.