How DataWorks Data Agent Advances from Augmented Assistance to Full Autonomy

At the 2026‑04‑29 Alibaba Cloud "Xia" event, senior technical expert Xu Ri introduced the latest upgrade of DataWorks Data Agent, describing it not merely as a Copilot enhancement but as a shift from "augmented" assistance to "full autonomous" operation.

1. From Pain Points to Paradigm Shift

The author outlines five evolutionary stages of data agents:

Stage 1 – Code Completion: The system suggests the next line of code after each keystroke.

Stage 2 – Q&A & Code Assistance: Natural‑language queries generate explanations and code snippets, enabling a "Ctrl C + Ctrl V" style development.

Stage 3 – IDE Copilot: The agent understands comments, translates code, and can reduce coding effort by 30‑40 %.

Stage 4 – Chat BI: Business users obtain data or simple reports via chat, though accuracy remains a concern.

Stage 5 – Autonomous Data Agent (current release): The agent performs end‑to‑end tasks—from requirement understanding, data exploration, code generation, to job deployment and post‑deployment attribution—without human intervention beyond a single confirmation.

2. Dual‑Engine Architecture: CLI and Claw Modes

DataWorks Data Agent employs two complementary modes that share a unified context:

CLI Mode: Optimized for complex code‑centric tasks. It reads project files, change logs, performs data insight, generates code, unit tests, quality rules, and hands the result to a review step before publishing.

Claw Mode ("龙虾"): Designed for incident‑driven scenarios. Integrated with DingTalk, WeChat Work, Feishu, etc., it autonomously diagnoses alerts, proposes actions, and executes them after a simple user confirmation.

Both modes access the same semantic context, enabling seamless hand‑off between development and operations workflows.

3. Unified Technical Kernel

The platform builds a true unified runtime rather than merely embedding a generic Code Agent and a Claw engine. At the top sits an ACP gateway that routes user intents to the appropriate agent type (Code or Claw). For example, a request to split a table is dispatched to the Code Agent, whereas a diagnostic query about a failed job is routed to the Claw Agent, with results delivered via chat.

The shared kernel encompasses model, container, execution engine, encryption, permission, and observability layers, ensuring consistent security and governance across modes.

4. Open Ecosystem & Fully Managed Runtime

DataWorks Data Agent is built on the existing DataWorks infrastructure, reusing resource groups, cloud‑native runtimes, workspace bindings, and permission systems for zero‑cost cold‑start. It supports multiple interaction surfaces—CLI, TUI, chat UI, IDE plugins, and APIs.

An open MCP‑based Skill ecosystem allows partners and customers to extend the agent’s capabilities. Once a Skill is registered, it becomes universally available across scenarios. The platform also integrates major Chinese LLMs (Tongyi Qianwen, GLM, DeepSeek) and offers fine‑tuned Text‑to‑SQL models for big‑data workloads, with the option to deploy private models.

5. CPU‑GPU Joint Optimization Insights

While many assume GPU acceleration alone drives agent performance, the article highlights that CPU consumption in tool‑side processing significantly impacts latency. Alibaba’s DataWorks team collaborated with AMD and Intel to optimize physical core frequency and thread throughput, improving overall agent execution efficiency.

6. Conclusion – A New Starting Point

The release marks a paradigm shift from "enhanced" assistance to "autonomous" operation. By combining CLI and Claw engines, a unified runtime, and an extensible Skill ecosystem, DataWorks Data Agent enables a single goal to trigger an end‑to‑end data pipeline. The author emphasizes that this is not an endpoint but a foundation for future digital‑employee capabilities in enterprise big‑data environments.