DataWorks Unveils Agent-Ready Lake & ADA for Enterprise Data Agents

Two Simultaneous Shifts: From Human-Centric to Agent-Centric Data Platforms

DataWorks product lead Tian Qixi stated that enterprise data platforms are undergoing two simultaneous shifts: data consumers are changing from humans to AI Agents, and data producers are changing from humans to Data Agents. The platform's end state is becoming Agent-Ready — built for AI Agents, powered by Data Agents. DataWorks now positions itself as an "Agentic AI era enterprise-grade data agent platform" to upgrade enterprise lakehouses into Agent-Ready Lakes.

DataWorks' 16-Year Evolution to Agent Native

DataWorks began in 2009 as a scheduling engine, evolved through data integration, a one-stop governance platform, and the 2024 Copilot assistant, reaching full Agent Native in 2026. It handles 40M+ task instances daily and manages EB-scale data, providing a massive production training ground for Agents. In IDC's 2026 China Data Agent Vendor Assessment, DataWorks placed first in the Leaders quadrant for both capability and strategy. It has also held the top market share in China's data governance platform market for five consecutive years and supported Alibaba Cloud's Leaders position in Forrester's 2025 Q4 AI Platforms and AI Infrastructure Solutions evaluations.

Tian acknowledged a key reality: building an impressive Agent demo is easy, but deploying Agents into real enterprise data production lines is far harder. Agents can write syntactically correct SQL but often fail semantically — they may not find the right tables or metrics. Worse, humans cannot easily detect when an Agent's conclusion is wrong. "Today it's not that models aren't smart enough; it's that Agents lack business common sense."

Context Graph: The Business Brain for Data Agents

To solve the "no business common sense" problem, DataWorks introduced the enterprise-grade semantic graph Context Graph. It acts as the Data Agent's business brain — letting Agents understand business before acting. Construction uses AI to automatically scan and reason over existing platform assets: metadata, scheduled jobs, metric definitions, data lineage. The semantic Agent identifies entities and relationships, infers metric definitions, and annotates sources with minimal human intervention, supporting incremental updates as assets change.

Context Graph provides six semantic layers: metadata enrichment, metric terminology, relationship & lineage, scheduling & quality, intelligent learning, and organizational knowledge — building a complete, fully traceable semantic system from field meanings to organizational knowledge.

With Context Graph, internal tests show business Q&A accuracy of 93.24% and SQL generation consistency of 99% . Tian emphasized consistency: "Ensuring the same question gets a stable, consistent answer across the enterprise, reducing model hallucination, is critical." Adding the semantic layer allows a Flash-model-class LLM to achieve near-flagship-model performance, drastically cutting token costs.

"If past DataWorks governance was for humans managing data, today Context Graph is governance for Agents" — it will determine future Agent effectiveness. DataWorks plans to open semantic retrieval via MCP services so any enterprise Agent can access the warehouse through semantic retrieval.

Full-Process Expert Suites: 15 Official Suites, 80+ Skills

Agents need skills beyond business understanding. DataWorks releases expert suites covering nearly all core modules: 15 official suites with 80+ professional skills across data integration, development, operations, governance, and analysis — ready out of the box. "Data Agent is not a smart chatbot; it's a digital employee that can do work."

On scheduling, DataWorks moves Agent task scheduling to the cloud with managed scheduling. Even if an Agent sandbox instance stops due to idleness, scheduled orchestrations continue running 7×24 in the cloud, with results pushed via DingTalk, Feishu, or WeCom. Users can wake a sleeping Agent instance directly from the IM notification. This mechanism greatly reduces sandbox instance costs.

Business process capabilities let enterprises codify fixed workflows into standard processes for organization-wide sharing and reuse.

Three-Layer Security Governance

Stronger Agents mean higher risks. DataWorks builds three defense lines:

Agent Data Security: Agent identity management, fine-grained data permissions, cloud-hosted secure sandbox — ensuring data never leaves the domain.

Agent Behavior Audit: Full-chain observability and audit; extends data lineage to Agent task lineage — showing which person used which Agent, which session generated which task, which table was produced. Includes data watermark tracing and risk assessment of behaviors.

Agent Cost Control: Credit quota management and Token cost governance for Agents, preventing cost overruns.

Open Extensibility and Autonomous Inspection

DataWorks Data Agent offers six open capabilities:

Agent engine: built on Qwen Code, also supports OpenClaw, Hermes Agent, DeepSeek Harness, QwenPAW for custom Agents.

Semantic graph: open data ingestion and export.

Expert suites, skills, MCP connectors: all support enterprise customization.

OpenAPI: embed Data Agent into enterprise systems for bidirectional integration flexibility.

On autonomy, DataWorks launches autonomous inspection service. Agents no longer wait for instructions; they proactively discover, diagnose, and remediate. Triggered by schedule or events, Agents auto-detect issues, locate root causes via logs, propose and execute fixes, and produce daily inspection reports. Coverage currently spans data integration, scheduling operations, and data governance.

Multi-Cloud, Multi-Modal, Multi-Engine Development on OpenLake

Around the OpenLake architecture, DataWorks upgrades three data development capabilities:

Unified multi-cloud data integration: Supports 63 heterogeneous data sources across Alibaba Cloud, AWS, GCP, Azure, Tencent Cloud, Volcano Engine. Real-time and offline ingestion. New multi-modal processing: a single ETL job parses images, audio/video, documents during ingestion — data lands processed, no post-processing chain. Natural-language configuration boosts efficiency by 80% .

Unified multi-cloud multi-modal metadata management: Dedicated data catalog for cross-cloud, multi-tenant, multi-source metadata centralization; data stays in place. Manages structured data, raw files, multi-modal datasets. Upgraded Agent data lineage shows full process of Agent involvement in data processing.

Unified heterogeneous multi-cloud task scheduling & Data+AI development platform: Single scheduler for both periodic big-data loads and trigger-based AI loads. Three interaction paradigms: smart IDE, Notebook, Data Agent. Supports AI Function, Python, PySpark, Ray compute paradigms for multi-modal data pipelines.

Data Agent Architecture and Live Commerce Demo

DataWorks Data Agent R&D lead Zhu Jiadun demonstrated a live-commerce operations scenario. Analyst Chen receives a 9 AM request: "Check last month's live-stream promotion new-user conversion; need it for 2 PM meeting." Traditional engineering would take two days. Chen uses one sentence to trigger Data Agent, which auto-locates tables, confirms metrics, traces lineage, queries and analyzes. On finding anomalies, it connects to DingTalk group chat, analyzes live-stream video, compares promotion creatives — achieving multi-modal fusion analysis. Chen delivers a full analysis report in ~30 minutes.

The demo reveals the Managed Data Agent architecture: user interaction layer (Web, DingTalk, Feishu, WeCom), platform control layer (full lifecycle management, security audit), Agent runtime layer (based on Qwen Code, extended via MCP and Skills). Autonomy levels follow self-driving standards L1–L5; most scenarios reach L3 (conditional autonomy) , some reach L4 (scoped autonomous delivery) .

ADA: AI-Native Big Data Service for Multi-Agent Collaboration

Building on DataWorks' five capability upgrades, Alibaba Cloud introduces ADA (Agentic Data and AI Platform). ADA fuses agent technology with big data and AI engines to provide a next-gen data productivity platform, enabling natural-language big data tasks and accelerating data decisions. In ADA, Agents from MaxCompute, Flink, EMR, etc., achieve multi-agent collaboration.

ADA's three core features:

Multi-agent collaboration: AdaWork acts as primary Agent for auto-coordination and routing, completing cross-engine complex delivery tasks.

Knowledge engine: Same Context Graph semantic graph capability as DataWorks.

Custom Agent capability: Low-code and high-code options; enterprises can choose mainstream Agent engines to build proprietary agents.