Tencent's RecSys Factory bounds LLM agent autonomy to decision points

RecSys Factory is a new LLM-agent platform from Tencent researchers (Dongyang Ao, Kaixiang Fang, and Shijie Xu) designed to solve what they call the autonomy-determinism-efficiency trilemma in industrial recommender systems. The core idea: don't let agents roam freely across entire pipelines. Instead, grant autonomy only at discrete decision points, where operators need interpretation, schema-conforming feature extraction, or glue-code generation. This bounds risk while preserving flexibility. The paper reports a 78-day deployment across three heterogeneous Tencent recommender business lines, covering 1,624 CLI-tool dispatches with a 78.6% aggregate success rate.

The architecture has three deconstructions. First, runtime is split into event sources—Claude Code Stop hooks, corporate-IM webhooks, and workflow scheduler APIs—so the platform runs no daemon during wait phases and uses zero CPU during the 94% of wall-clock time spent waiting on Spark or GPU jobs. Second, capability is broken into a 29-file skill ecosystem (8,971 lines) whose per-skill pitfall tables compile into a 400-entry PitfallStore, confining autonomy to typed decision surfaces inside pre-committed pipelines. Third, humans stay in the loop via a schema-validated, idempotent, replayable card protocol at the diagnostic-versus-execution boundary, tested in an 8-day, 16-run pilot. The authors frame performance observations as case studies, not controlled claims, but the results suggest a path to safe LLM agent deployment in high-stakes industrial settings.