Huawei Cloud AgentArts / Pangu Agent

1. Introduction

1.1. Platform Positioning

AgentArts (ZhiGuo) is Huawei Cloud's agent platform for government/enterprise and industrial-manufacturing scenarios, and is the productized carrier of Pangu Foundation Model's agent capabilities. Its positioning can be summarized in three keywords: Ascend ecosystem (full-stack domestic-indigenous models and compute), government/enterprise agent (full-stack Xinchuang adaptation), and deep vertical focus on industrial manufacturing (equipment/production-line integration and device-edge-cloud collaboration).

In March 2026, the Huawei Cloud Partner Conference announced a new AI layout: centered on industry agents, with a newly open-sourced Agentic foundation-model pool (Pangu open-sourced at full scale + 160+ aggregated models including DeepSeek and Qwen); the enterprise commercial edition entered public beta in April 2026. Pangu agent capabilities are delivered in productized form through AgentArts, and this article writes the two as a single platform (consistent with the treatment in research report AG-6).

Public-availability note: AgentArts' official documentation and architecture materials have low public availability, so this article marks many items as [to be verified] / [to be filled in] (consistent with the group-wide gap in section 6.4 of the group README). The ratings reflect the density of publicly verifiable mechanisms, not the ceiling of its capabilities.

1.2. Basic Information Card

1.3. Development Timeline

1.4. Position in the AI Harness System

Per the boundary definitions of the parameter card, AgentArts is the government/enterprise Xinchuang form of Agent Platform (Harness' productized packaging), and together with HiAgent (see 19-hiagent.md) it forms the domestic twin at the "government/enterprise privatized workstation" lineage slot:

• Compared with HiAgent: HiAgent focuses on the ByteDance engineering ecosystem (Feishu / Doubao) and cluster management, while AgentArts focuses on Xinchuang certifications (Kunpeng / Ascend / openEuler, MLPS Level 3) and industrial depth (device-edge-cloud, production-line integration);
• Compared with Alibaba Cloud Bailian (14-alibaba-bailian.md): Bailian is primarily a public-cloud SaaS + open-source base, while AgentArts is primarily government/enterprise privatization + an open-source model base (Pangu open source);
• Compared with framework-type platforms such as 03-google-adk.md: AgentArts is not a framework but a platform; orchestration details have low public availability, so the six-layer ratings are based on platform-level publicly available capabilities;
• Pangu Agent capabilities are not covered in a separate article: their productized export is AgentArts (consistent with the exclusion logic of the research report).

2. Glossary

3. Functional Description

3.1. Low-Code Building and Industry Templates

• Low-code visual orchestration: business and implementation staff build agents by drag-and-drop, aimed at non-algorithm engineers;
• Industry templates: agent templates preset by industry scenario (manufacturing inspection, process optimization, etc.) to shorten delivery cycles;
• The specific list of industries covered by the template library is not fully public, marked [to be verified].

3.2. Open-Source Agentic Foundation-Model Pool and Semantic Routing

• Open-source model pool: Pangu open-sourced at full scale + 160+ aggregated models including DeepSeek / Qwen, letting customers freely choose within the domestic model ecosystem;
• Semantic routing (ModelArts upgrade in 2026-06): automatically routes to the best-matching model by task semantics — the key mechanism that upgrades the model pool from "choose one of many" to "intelligent scheduling," and a differentiator at the L1 / model layer;
• Third-party model integration is fast (GLM-5.1 integrated on the day of launch; Kimi / Zhipu / DeepSeek integrated into ModelArts).

3.3. Device-Edge-Cloud Collaboration and Production-Line Integration

• Device side: Pangu 1-5B small models can be deployed locally on edge devices, covering offline / low-bandwidth workshop scenarios;
• Edge: agents can be deployed on edge nodes, connecting directly to production-line equipment, with edge data encryption protection;
• Cloud side: ModelArts model matrix + semantic routing + reinforcement learning service handle training and complex inference;
• Equipment/production-line integration: the ability to interface with SCADA-type production-line systems is what anchors it in industrial vertical scenarios (details not public, [to be verified]).

3.4. Xinchuang Compliance System

• Full-stack domestic-indigenous: Kunpeng (general compute) + Ascend (AI compute) + openEuler (OS) + Pangu (models);
• Certifications: MLPS Level 3 + Xinchuang certification;
• Data security: industrial data security protection and edge encryption, data never leaving the domain;
• This system is AgentArts' core competitiveness in strictly compliant scenarios such as central state-owned enterprises, defense, and government.

4. Platform Architecture

4.1. Three-Layer Device-Edge-Cloud Carriage

1. Cloud layer: ModelArts (model matrix, semantic routing, reinforcement learning service) + AgentArts platform services (orchestration, templates, governance) + 160+ model pool;
2. Edge layer: edge nodes host the agent runtime, connecting directly to production-line equipment and industrial systems, with edge encryption;
3. Device layer: Pangu 1-5B edge models deployed locally, performing lightweight inference and on-site perception;
4. Cross-cutting layer: Xinchuang foundation (Kunpeng / Ascend / openEuler) + MLPS Level 3 security system covering all three layers.

4.2. Execution Flow of an Industrial Quality Inspection Agent Task

1. The inspection agent is instantiated via an industry template and deployed to a factory edge node;
2. Production-line images are captured by equipment and enter edge inference (Pangu multimodal / edge small models split the work);
3. Difficult samples are semantically routed to larger cloud-side models for review;
4. Detection results are written back to production-line systems (MES / SCADA type), and anomalies trigger handling workflows;
5. Data is edge-encrypted throughout and never leaves the factory domain (cloud review, if needed, goes through a secure channel);
6. Inspection knowledge and misjudgment cases are accumulated as enterprise knowledge for later iteration (mechanism details [to be verified]).

5. Harness Design

5.1. Six-Layer Capability Overview

Strength/weakness assessment: L6 is the absolute strong point (the deepest Xinchuang certification system in this group), L1 is next (model-pool semantic routing), and L3 / L4 / L5 have insufficient public availability, with "Medium" used as the conservative baseline.

5.2. L1 Context Engineering Layer

• Semantic routing is the most evidence-supported mechanism in this layer: task semantics decide which model handles the job, essentially automating "context → model selection," a cross-innovation between L1 and the model layer;
• Industry knowledge fusion: Pangu industry models (manufacturing, mining, meteorology, etc.) carry built-in domain knowledge injection;
• The 160+ model pool ensures model-side elasticity for context processing;
• Public details of RAG / retrieval mechanisms and context compression are missing, marked [to be verified].

5.3. L2 Tools & Execution Layer

• Edge-node deployment is this layer's differentiator: the agent runtime sinks to factories / parks, satisfying execution latency and data residency at once;
• Equipment/production-line integration: data interfacing with industrial systems (production-line equipment, inspection instruments) is a hard requirement of vertical scenarios;
• Edge-side 1-5B models provide offline execution capability;
• Tool protocols (e.g., whether MCP is supported) and sandbox mechanisms are not public, marked [to be verified].

5.4. L3 Orchestration & Control Layer

• Low-code visual orchestration + industry templates are the visible orchestration form, aimed at implementation staff;
• Primitives such as multi-agent collaboration, sub-agent dispatch, and interrupt recovery have no public evidence, marked [to be verified];
• Compared with HiAgent's cluster-management narrative: AgentArts' orchestration story leans more toward "single agent + industry solutions," with no public account of cluster-level management.

5.5. L4 Memory & State Layer

• No systematic memory mechanism is described in public materials, marked [to be filled in];
• The actual needs of industrial scenarios (accumulated inspection knowledge, equipment-ledger memory, misjudgment-case libraries) are expected to exist in the form of industry solutions, but product-level mechanisms await verification in official documentation.

5.6. L5 Evaluation & Observability Layer

• ModelArts' reinforcement learning service (2026-06 upgrade) provides training-side evaluation-reward-loop infrastructure;
• Runtime-side trajectory tracing, metric systems, and regression mechanisms of agents have no public description, marked [to be verified];
• Pangu's industry benchmark results (e.g., inspection accuracy) appear in scattered promotional accounts, lacking independently verified data.

5.7. L6 Governance & Security Layer

• Xinchuang system: Kunpeng / Ascend / openEuler full-stack domestic-indigenous — the deepest L6 certification depth among the 21 platforms in this group;
• MLPS Level 3 + Xinchuang certification: entry-level certifications for strictly compliant industries;
• Data never leaving the domain: triple protection of edge deployment + edge encryption + privatized delivery;
• Industrial data security protection: covers production-line data collection, transmission, and storage;
• This layer is rated "Strong" with the highest evidence density — AgentArts' product focus is clearly on the governance side.

5.8. Device-Edge-Cloud Collaboration Architecture Diagram

Figure 20-1 | AgentArts Device-Edge-Cloud Collaboration and Xinchuang Foundation

Data source: Huawei Cloud official accounts and public reports (as of 2026-09-12), illustrative.

6. Practical Cases

Case 1: Industrial deployment of the Pangu Manufacturing Model (official account)

The Pangu Manufacturing Model has been commercially deployed in industrial quality-inspection and process-optimization scenarios, serving as Huawei's flagship narrative for agents in manufacturing. No publicly disclosed, verifiable engineering metrics for specific customers — such as inspection-accuracy improvement or labor savings — were found.

Case 2: Aggregation of the model ecosystem (multiple accounts)

GLM-5.1 was integrated into AgentArts on the day it launched on Huawei Cloud; in 2026-06 ModelArts integrated Kimi / Zhipu / DeepSeek and introduced semantic routing and a reinforcement learning service — reflecting the actual pace of progress along its "open-source model pool + semantic routing" route.

Sections with no publicly available quantitative data: no public quantitative disclosures were found for the GA status after the enterprise commercial edition (public beta in 2026-04), customer lists, deployment scale, or horizontal comparison data with competitors. This is plainly noted here rather than filled in at will.

7. Summary

7.1. Strengths

1. Deepest L6 Xinchuang depth: Kunpeng / Ascend / openEuler / MLPS Level 3 / Xinchuang certification — the strongest admission standing for strictly compliant industries;
2. Complete device-edge-cloud coverage: the layered architecture of edge small models, edge runtime, and cloud-side model pool fits the industrial physical world;
3. Semantic routing: intelligent scheduling over a 160+ model pool is a differentiator at the model layer;
4. Pangu industry models: knowledge injection for verticals like manufacturing has long-standing accumulation;
5. Open-source model foundation: Pangu open-sourced at full scale reduces lock-in concerns;
6. Unique domestic ecosystem position: no equivalent competitors exist in domestic-substitution and Xinchuang-procurement scenarios.

7.2. Weaknesses

1. Low public availability: architecture, memory, evaluation, and orchestration details are not public, so the capability ceiling cannot be independently verified;
2. Insufficient evidence for L3 / L4 / L5: orchestration primitives, memory mechanisms, and evaluation systems have no public description;
3. Commercial-edition maturity: public beta only began in 2026-04; GA status and pricing are unclear;
4. Leaner orchestration narrative: cluster-level multi-agent management lags HiAgent;
5. Relatively closed ecosystem: migration paths outside the Huawei stack (Ascend / openEuler) are unclear.

7.3. Applicability Boundaries

7.4. Selection Recommendations

• Strictly compliant + industrial scenarios: AgentArts is the top candidate in this group; when comparing with HiAgent, decide between "certification depth vs. engineering ecosystem";
• If you need multi-agent cluster management: HiAgent's public account is stronger (see 19-hiagent.md);
• For lightweight public-cloud needs: Bailian is more suitable (see 14-alibaba-bailian.md);
• Before adoption, request architecture documents, evaluation methods, and GA / pricing explanations from the vendor, and verify each [to be verified] item in this article one by one.

Information-Gap Statement

1. Official documentation availability: AgentArts' architecture and six-layer details have low official public availability, so much of this article relies on [to be verified] items.
2. GA status and pricing: the GA status and pricing after the enterprise commercial edition (public beta in 2026-04) were not found, marked [to be filled in].
3. Pangu model license: the specific license type for Pangu's open source was not verified, marked [to be verified].
4. L3 orchestration details: no public evidence for multi-agent collaboration and control primitives, marked [to be verified].
5. L4 memory mechanism: no public description, marked [to be filled in].
6. L5 evaluation system: runtime-side evaluation and observability have no public description, marked [to be verified].
7. Customer engineering metrics: no publicly disclosed data with verifiable evidence for industrial deployment cases; not filled in.

8. References

1. Huawei Cloud — official website. https://www.huaweicloud.com/
2. ModelArts — Huawei Cloud AI development platform product page. https://www.huaweicloud.com/product/modelarts.html
3. Huawei Cloud Enterprise Business — official website. https://e.huawei.com/
4. Ascend — Huawei AI compute ecosystem official website. https://www.hiascend.com/
5. openEuler — open-source operating system community official website. https://www.openeuler.org/
6. Modelers community (魔乐) — domestic open-source model community. https://modelers.cn/
7. Report on Huawei Cloud AI layout and agent developments — China Economic News Network, 2026. https://www.cet.com.cn/xwsd/10523310.shtml
8. Industry news on agent platforms — Tencent News, 2026-08-22. https://news.qq.com/rain/a/20260822A02HE800
9. R18-IDE-Agents-supplementary platforms research report — internal research report of this project (AG-6 key fact card).
10. Project parameter card v1.1 (six-layer capability model and lineage boundaries) — internal baseline document of this project.