How Tencent Redefines Data Architecture for the Agent Era

When agents evolve from simple question‑answering to full execution, the conventional stack—data residing in a lake, models hosted in the cloud, and scheduling spread across two locations—reveals three fatal shortcomings: each training run requires moving petabytes of data; runtime trajectories, inference chains, and preference feedback are scattered in logs and cannot be fed back; and a split compute stack (Spark for ETL, Ray for training) fragments resources and doubles costs.

Tencent Cloud’s Big Data AI DLC addresses these issues by extending the data lake with both Spark and Ray engines that share identical metadata, permissions, and storage, allowing computation to occur directly where the data lives. Agent preprocessing, fine‑tuning, and inference are all performed on a single platform, creating a closed‑loop workflow.

The platform automatically captures execution traces and transforms them into the next round of training data, eliminating the need for data movement while continuously iterating knowledge. This integration of AI capabilities into the data lake redefines data architecture for the Agent era, merging what were previously separate data and AI platforms into a unified system.