How to Make DataAgent Truly Usable: Optimizing Enterprise Credit with Enhanced Metadata and Business Cognition

DataAgent must simultaneously solve two fundamental problems—accurate data acquisition and deep business cognition—to be truly usable in enterprise credit contexts. The Ant Credit team addresses three dimensions of complexity (users, scenarios, assets) by strengthening metadata, constructing a hybrid semantic layer, and building an evolvable cognition system, all driven by Harness Engineering for continuous iteration.

Recent advances in Coding Agent demonstrate that agents can reliably enter complex engineering workflows, not merely generate code. Improvements in model reasoning, long‑chain planning, and tool invocation, together with mature frameworks such as Claude Code, create a stable engineering loop that combines model capability with system orchestration.

Transferring this closed‑loop capability to data analysis requires dynamic context loading and unloading rather than a static prompt. DataAgent must manage a mutable set of work contexts that evolve with each task, ensuring that intermediate hypotheses and results do not accumulate unchecked noise.

The tool ecosystem is equally decisive: an agent limited to SQL execution is merely an automated query tool. DataAgent must also perform data validation, semantic recall, result verification, and knowledge loading to achieve stable analytical performance.

Enterprises need a composable set of DataAgent capabilities—exposed as Skills—rather than a monolithic product. These capabilities should be embeddable in digital assistants, workflow engines, or other business systems, enabling end‑to‑end analysis chains instead of isolated Q&A interfaces.

The proposed architecture consists of four layers: (1) a semantic layer that resolves "can we fetch the right data" through metadata enhancement, semantic models, metric and tag management; (2) a cognition layer that interprets business terminology, scenarios, and user preferences; (3) a core process layer that orchestrates planning, analysis, and reflective iteration; and (4) a channel layer (CLI, MCP, API) that determines how agents are consumed, either as Skills or standalone entry points.

Four key practices underpin the implementation: (1) Metadata enhancement to disambiguate thousands of tables and fields via profiling, confidence labeling, and expert review, exposing semantic services through MCP; (2) Mixed semantic solutions with three routing paths—NL→DAL→SQL for Chat BI, NL→Model→SQL using MetricFlow/Cube.dev for core insights, and NL→SQL with table metadata for exploratory queries—selected by a routing strategy; (3) Business cognition built from explicit knowledge extraction, manual curation, and agent feedback, organized into a chunk‑based vector store and a document repository that are dynamically loaded during analysis; (4) Leveraging Coding Agent to iteratively improve DataAgent, employing multiple specialized agents (Main, Reviewer, Executor, Eval) in a human‑in‑the‑loop workflow that aims to evolve toward human‑on‑the‑loop autonomy.

Looking ahead, the platform should shift from "human‑centric" to "agent‑centric" interfaces, support decentralized asset management, and adopt intelligent discovery combined with human‑machine collaborative governance. Continued investment will focus on semantic data assets, deeper business cognition, and self‑evolving agent capabilities to close the decision loop in AI‑driven enterprises.