Kuaishou Demonstrates DataAgent: A New Paradigm for Data Intelligence

Han Jiang, R&D Lead of Kuaishou’s Production Platform R&D Center, has confirmed participation in QCon Global Software Development Conference Shanghai 2026, scheduled for October 22-24. He will present “DataAgent: Kuaishou’s Intelligent Exploration of Big Data Production and Analysis”. DataAgent is an enterprise-grade Agent designed for data domains with general-purpose capacity, already achieving nearly ten thousand weekly active users in proactive conversations, supporting core business scenarios including data production, analysis, and A/B testing.

Key facts:

• Launch timeline: Officially disclosed at QCon Shanghai 2026 (Oct 22-24)
• Technical position: Enterprise-grade data intelligent agent covering the full data lifecycle
• Core capabilities: data discovery, development, root-cause reporting, audience insight, A/B analysis, behavioral event analysis
• Adoption metric: nearly 10,000 weekly active users (proactive sessions)
• Architecture: Multi-Agent (parent-child Agent + Skill) coordination framework
• Leadership: Han Jiang, who led Kuaishou’s “Tiangong” big data platform from 0 to 1, with 10+ years of experience

Multi-Layer Architecture: Tackling Dual Challenges in Data Workflows

DataAgent addresses two layers of complexity: universal Agent issues like Multi-Agent coordination, context management, and execution reliability, plus data-domain specifics including large-scale knowledge bases, data heterogeneity, accuracy maintenance, and strict security requirements.

Parent-Child Agent + Skill Architecture

The parent-child Agent design coordinates general versus domain-specific capabilities, with Skill components enabling dynamic functional extension. Space isolation and isolation packages support business-line segmentation needs.

Intelligent Context Management

Dynamic loading of Skills & Tools, intelligent context unloading/compression, and intelligent short/long-term memory mechanisms enable continuous learning—“smarter with use.” This manages query complexity across hundreds of thousands of data assets while preventing context explosion.

Knowledge Engineering: The Self-Evolving “Knowledge Flywheel”

Kuaishou moves beyond traditional document-centric knowledge bases with LLM-wiki, organizing knowledge as “human-AI co-built, structurally unified, versionable assets directly consumable by LLMs.”

The knowledge flywheel forms a closed loop:

• Query → Retrieval → Feedback → Attribution → Iteration → Knowledge Update → Validation

This ensures knowledge remains “fresh and active,” countering knowledge decay—a key differentiator from static documentation systems.

Validation and Real-World Impact: From Tech to Business Value

DataAgent has been deployed for end-to-end data production and intelligent analysis at Kuaishou. For initial rollout, a seed dataset + precision standard + self-growing soil strategy ensured high-quality knowledge entry.

Critical implementation metrics:

• Supported scenarios: data production, analysis, A/B testing
• User base: ~10,000 weekly active users (proactive sessions)
• Scale handled: retrieval and disambiguation at hundreds of thousands of asset scale

Practical Recommendations for Developers and Enterprises

Target audiences should consider:

• Data platform teams building intelligent agents: Adopt Multi-Agent coordination and knowledge flywheel mechanisms
• AI Infra engineers facing context explosion and reliability: Study dynamic loading and compression strategies
• Business decision-makers planning “data-driven → insight-driven” transition: Review adoption proof (~10k weekly users demonstrates feasibility)

Teams in early exploration phases should pilot single-scenario implementations while planning feedback loops early—knowledge without real-world problem feedback cannot sustain accuracy.

Final Thoughts

DataAgent demonstrates that AI for Data is not about adding LLMs to existing workflows, but about paradigm shift through mode reconstruction. As agents evolve from tools to workflow orchestrators, the focus shifts from data discovery to problem understanding—marking the maturation of intelligent data systems.