How 6 Chinese Tech Giants Engineer Context for Production AI Agents

Overview

The article previews six talks from DACon 2026 in Beijing, each detailing how a major technology company tackles context engineering — the discipline of organizing, compressing, retrieving, and keeping fresh the information that AI agents need to operate reliably in production. The common thesis: agent failures (repetitive questions, hallucinations, forgotten corrections) stem not from model capability but from poor context quality.

01 | JD.com: Multi-Layer Dynamic Memory

Speaker

Gu Licheng, Algorithm Engineer at JD Retail.

Core Problem

B2B agents require long-term, production-grade, reusable context. The barrier is not agent count or orchestration complexity but context management quality.

Solution: Three-Layer Memory Architecture

Long-term memory : Stores user profiles and cross-session preferences/focus. Handles extraction, update, and ranking.

Mid-term memory : Stores multi-turn conversations. Manages intent continuity, intent switching, compression, and rollback.

Short-term memory : Stores task execution process. Handles observation processing and complex information simplification.

Engineering implementation covers four dimensions: storage scheme, retrieval strategy, injection timing, and window management.

Production Challenges

Context isolation and merging during DAG orchestration.

Context reuse during persistence.

Step-level retry enabling agent self-correction of queries.

Human-in-the-loop intervention: context survival and recovery when re-running from step N.

Key Takeaways

Real-world experience with DAG orchestration, human intervention, and partial re-runs — beyond demo-stage memory design.

Hardest Hurdle

Defining measurable metrics for memory effectiveness (recall accuracy, injection correctness) to replace gut-feel optimization.

02 | Hangzhou Qunhe: From Workflows to Agent Clusters

Speaker

He Bihong, SRE Team Lead & Software Development Expert at Hangzhou Qunhe.

Evolution

Started with AI workflows for AIOps (predefined nodes for fixed processes). Business evolution exposed bottlenecks: no dynamic decision-making, no persistent memory, no multi-role collaboration. Switched to a self-developed multi-tenant digital employee system on the HermesAgent base, now scaled across SRE AIOps, marketing, customer service, DevOps.

Three Context-Related Pillars

Memory & Chain-of-Thought Coupling : Bind long-term business memory with standardized CoT. Rule-based memory constrains model output for stable, traceable reasoning. Standardized CoT also compresses reasoning paths, removing redundant thoughts for precision and performance.

Self-Evolving Flywheel : Business operations precipitate experience; experience iterates the knowledge base. Supports dedicated, business, enterprise stock knowledge, and knowledge graphs in multiple modalities.

Isolation & Governance : Three-layer isolation (data, session, memory) for multi-tenancy. SSO integration for unified token authentication. Skills follow unified auth specs with clear boundaries and mutex rules.

Collaboration Model

Abandoned unordered multi-agent negotiation in favor of process + state machine : each digital employee binds to a dedicated role and task state; standardized state transitions and task handovers drive complex business closure. Automated in incident handling, repair, code submission — shifting ops capability right.

Core Conclusion

Enterprise-grade agent moats are engineering, governance, security, and collaboration — not raw model capability.

Hardest Hurdles

Continuous capability precipitation and evolution: skill sprawl, scattered experience, difficulty maintaining efficiency.

Performance: high latency from LLM and native agent reasoning chains; uncontrolled redundancy fails millisecond-level business SLAs.

03 | Guanyuan Data: Decision Context, Evidence Chains & Evaluation Loops

Speaker

Shi Kai, Product Expert at Guanyuan Data.

Premise

Enterprises already possess massive data assets and decision knowledge. The challenge is converting these into agent-understandable business context for operational decisions — not rebuilding from scratch.

Two High-Value Sections

Decision Context Engineering : How to organize and apply data assets and decision knowledge. Context compilation logic follows business questions, not tables.

Trusted Evidence Chains : Link data, analysis, and conclusions so that analytical results are explainable, traceable, decisions are evaluable, and action effects are auditable.

Customer Case Studies

Pain points, implementation paths, and practical reflections — including a continuous review mechanism from suggestion quality to business outcomes.

Hardest Hurdle

Context-business mismatch. Solution: dynamically trim context by problem and user permissions, then continuously update for accuracy and efficiency — context is not built once but sliced on-demand.

04 | Datastrato: Open Semantic Layer, Ontology & Metadata as Three Pillars

Speaker

Du Junping, Founder & CEO of Datastrato (Apache Gravitino initiator).

Foundational Judgment

As agents move from Q&A to reasoning, decision, and execution, data accessibility alone cannot support reliable intelligence. Agents must understand what data means , how business entities relate , and which rules constrain actions .

Three-Layer Pillars

Unified Metadata (base): Cross-cloud, cross-engine, multi-modal unified management via Apache Gravitino. Consistent description and access interfaces for data assets.

Open Semantic Layer : Share metrics, dimensions, business definitions across tools — eliminate "one metric, multiple definitions".

Ontology : Express business entities, relationships, and rules so agents map natural-language business intent precisely to real data and operational boundaries.

Build Path Analysis

Top-down (task-driven) : Fast value validation but creates duplicate modeling and isolated context silos.

Bottom-up (metadata-governance-driven) : Broad coverage, solid foundation, but technical structure alone cannot derive business meaning.

Hybrid (recommended) : Business scenarios drive modeling; unified metadata connects business concepts, data assets, and governance rules; business validation and agent usage feedback drive continuous iteration — turning scenario knowledge from "one-off models" into "shareable infrastructure". Productized as Datastrato 2.0 (Gravitino 2.0 Enterprise), evolving data catalog into AI Context Platform.

Hardest Hurdles

Context freshness: data drift, rule changes, model upgrades invalidate context; most teams lack automated validation and iteration mechanisms.

Collaboration cost: business meaning cannot be auto-derived from technical structure; requires deep business-side involvement in modeling.

05 | Li Auto: Semantic Layer as Context — Turning Expert Experience into Versioned Agent Assets

Speaker

Qian Han, Big Data Platform Lead at Li Auto.

Core Proposition

Production agent bottleneck is context quality: the difference between fragmented scraps in prompts/tools/logs vs. semantically governed, understandable, traceable structured context.

Approach: Spark/Flink Fault Diagnosis as Entry Point

Decompose knowledge scattered in prompts and tools into four layers with clear responsibility boundaries:

Ontology

OSI

Playbook

Generic Execution Layer

Agents perform cross-log, metric, data-warehouse forensics to produce evidence-based root-cause judgments , not guesses. Each diagnosis exposes gaps that are then precipitated into versioned, verifiable, regressible semantic assets — directly answering "how to keep context fresh".

Real Trade-offs Discussed

How far lightweight modeling must go to be sufficient.

How to add query constraints.

Balancing model exploration capability with engineering determinism.

Hardest Hurdles

Ontology, metric relationships, and playbooks still require heavy manual curation; cross-business replication efficiency is limited. Next step: standardize semantic assets and push auto-generation.

Full runtime control plane: permission control, execution tracing, evidence status, result verification, multi-environment adaptation.

06 | AWS: The Two Ends of the Action Loop

Speaker

Huang Da, Solutions Architect at AWS.

Thesis

While everyone optimizes the "middle" of Data Agents (perceive, analyze, decide, execute), the loop's success hinges on two neglected ends.

First End: Dare to Act — Amazon Bedrock AgentCore

Identity : Least-privilege identity for every action — authorized, auditable.

Gateway : Reliable, governed tool/API calls — solve "execution goes off rails".

Memory : Short- and long-term memory for cross-session recall.

Runtime : Session-level isolation and elastic scaling for safe production runs.

Observability + Evaluation-Driven Development (EDD) : Metrics to verify "how well it works".

Second End: See the Result — ThinkingAI Agentic Engine

Business users query data in natural language; real-time dashboards show "what is happening now", not "what happened yesterday". MCP calls Data Query, Dashboard Creation, Trigger Workflow — forming "AWS production base + ThinkingAI data capability" bidirectional complement. Extended to A/B experiment delivery to user agents, closing "observe → experiment → intervene" loop. Includes cross-industry (gaming, OTA, content) PoC-to-scale challenges and pitfalls.

Hardest Hurdle

Base-layer loss of control → "dare not let it actually execute". Missing any of: identity/permission control, reliable tool calls, memory, production isolation, evaluation — agents stay at "give advice", never reach execution.

Cross-Cutting Consensus

All six companies agree: context is not a one-time build . JD calls it "Memory Flywheel — make context more accurate with use, not more chaotic"; Qunhe calls it "self-evolving flywheel"; Li Auto calls it "precipitate diagnostic gaps into versioned, verifiable, regressible semantic assets"; Datastrato calls it "business validation and agent usage feedback continuous iteration". Different names, same principle: get it running first, then let the running process cultivate the context .

Selection Guide

Agent forgets after multi-turn, repeats old mistakes → Watch JD (three-layer memory) and Qunhe (memory governance & flywheel). JD explains layering; Qunhe explains multi-tenant safety.

Team stuck on "inconsistent metric definitions, agent answers wrong when tools change" → Watch Guanyuan (decision knowledge & evidence chains) and Datastrato (metadata + open semantic + ontology). Guanyuan gives business-decision perspective; Datastrato gives infrastructure standard. The top-down vs. bottom-up analysis alone can save a re-architecture cycle.

Agent can analyze but no one dares let it act → Watch Li Auto (evidence-based root cause with confidence and answer boundaries) and AWS (two-end solution checklist). Li Auto shows how to structure context for evidenced output; AWS lists concrete fixes for "dare not act" and "cannot see".