How Tencent Redefines Data Architecture for the Agent Era

When agents transition from simple question‑answering to full execution, three fundamental shortcomings appear in conventional data architectures: the data resides in a lake while models live in the cloud, scheduling is split across locations, and each training cycle requires moving petabytes of data.

The first issue forces massive data transfer before every training run. The second scatters agent execution traces, inference chains, and preference feedback across logs, preventing them from feeding back into subsequent training. The third uses separate engines—Spark for ETL and Ray for model training—resulting in fragmented compute resources and doubled costs.

Tencent Cloud’s big data AI DLC addresses these problems by extending the data lake with both Spark and Ray engines that share identical metadata, permissions, and storage. Computation occurs in place, allowing agents to perform preprocessing, fine‑tuning, and inference within a single platform. Execution trajectories are automatically persisted as training data for the next iteration, eliminating data movement while continuously enriching knowledge.

This integrated approach is more than a combination of a data platform and an AI platform; it embeds AI capabilities directly into the data lake, effectively redefining data architecture for the Agent era.