Evolution and Engineering Practices of DataWorks Data Agent

01 Cognition Shift: Three Stages of Data Agent Evolution

Since 2023, DataWorks has progressed through three incremental phases. The first phase, Copilot , offers SQL completion and generation as an assistive tool, improving coding efficiency by roughly 30%–35% . The second phase introduces human‑AI collaboration , where efficiency gains range from 30% to 100% as Agents gradually replace traditional SaaS GUIs. The third phase envisions AI‑autonomous operation: humans merely assign tasks, while AI orchestrates, executes, reviews, and decides, potentially delivering ten‑fold to hundred‑fold efficiency improvements.

02 Product Matrix: Four Agent Types Covering the Entire Data Chain

DataWorks Data Agent is not a single function but a layered service built on a model layer (Qwen series, GLM series, NL2SQL‑fine‑tuned expert models) and an agent layer offering four categories:

Data Engineering (ETL development)

Data Governance

Data Analysis (Chat BI)

Cluster Control & Operations Optimization

The interaction layer supports multiple UI forms, including Chat UI, CLI/Web terminal, remote‑control via QR code, and IM channels (DingTalk, Feishu, WeChat Work).

03 DataWorks Data Agent 2.0: Cloud‑Managed Engineering Practice

Earlier agents ran on personal machines, requiring 7 × 24 hours of manual work and facing security, risk, and compliance challenges. Data Agent 2.0 adopts a dual‑engine architecture based on QwenCode and OpenClaw , delivering a cloud‑sandbox that runs continuously ( 7 × 24 hours ) and integrates with enterprise production systems. Security is enforced through Alibaba Cloud’s Global Acceleration, PrivateZone , PrivateLink , and a dedicated DataClaw line, ensuring no data leaves the private network. All write operations require secondary identity verification.

The system provides four interaction modes:

Chat UI – natural‑language dialogue.

CLI/Web terminal – for developers and power users.

Remote‑control – scan a QR code to mirror the PC interface on a mobile device.

IM Channel – integrates with DingTalk, Feishu, and WeChat Work.

04 AI Assistant Service: Secure, Controllable Operations Assistant

Built on OpenClaw , the AI assistant addresses three enterprise‑level concerns:

Fully managed, no‑ops: One‑click instance launch provides 7 × 24 hours of online service without manual configuration.

Security & control: Private networking (PrivateZone, PrivateLink) and role‑based execution ensure all traffic stays within the corporate network; write actions require double‑confirmation.

Built‑in Skills: Pre‑packaged skills cover task diagnosis, workspace diagnosis, alarm analysis, task remediation, and quality monitoring.

When a task fails, the assistant pushes an alert to the IM channel, automatically performs root‑cause analysis, and can remediate (e.g., re‑run a task after updating an expired resource group) without opening a PC.

05 Case Study: Taobao Flash‑Sale Deployment

In Alibaba’s internal environment, the traditional IDE‑based data‑development workflow required hours to days per feature, suffered from low efficiency, inconsistent standards, and limited knowledge reuse. After switching to Data Agent, end‑to‑end intelligent development covered the full pipeline (ODS → DWD → DWS → ADS). By extending custom Skills and a business knowledge base, the development cycle shrank from 12–23 hours to 5–10 minutes . Automated Skill‑driven workflows enforced standards, systematic quality checks ensured data reliability, and accumulated knowledge was continuously reused.

The overall impact is a shift from manual, hour‑level development to minute‑level, autonomous governance, fundamentally changing the data platform’s value chain. Whether Data Agent is ready for large‑scale rollout depends on an enterprise’s willingness to inject norms, knowledge, and best practices into the agent’s evolution loop.

Conclusion : DataWorks Data Agent demonstrates how a data platform can evolve from assistive tools to fully autonomous AI agents, delivering dramatic efficiency gains, tighter security, and a unified, cloud‑native operational model.