Tencent Yuanqi / Yuanbao Agent Platform (Tencent)

1. Introduction

1.1. Two-Product Structure: Yuanqi and Yuanbao

"Tencent Agent" is carried by two products within the Tencent system, and they must be separated first during selection and evaluation:

The relationship between the two can be summarized as: Yuanqi is where you "build", and Yuanbao is where you "run" and "use"; agents created in Yuanqi can be published to Yuanbao with one click, and simultaneously to Official Accounts, QQ, Mini Programs, WeCom, and other channels.

1.2. Basic Information Card

All version numbers, model parameters, and pricing in this article are accurate as of the September 2026 retrieval point; Tencent's official technical documentation is sparsely public, so much information can only rely on third-party evaluations, and this is flagged item by item in the body text and the Information Gap Statement.

1.3. Development Timeline

1.4. Position in the AI Harness System

By the boundary table of the parameter card, Tencent Yuanqi belongs to the "channel / ecosystem-oriented" branch of Agent Platform (Harness's productized encapsulation). But compared with similar platforms such as Dify and Coze, its focus is clearly different:

• Dify's focus is the engineering loop (RAG depth, DSL, private deployment), Coze's focus is channel distribution + open-source engine, while Yuanqi's value centers on WeChat ecosystem distribution and the monetization loop — Official Account knowledge base auto-sync, one-click publishing, and WeChat Pay MCP form an end-to-end chain that other platforms cannot replicate;
• On runtime and engineering capabilities, there is very little public information: no architecture documents, no state-management documentation, no evaluation system. In the six-layer rating, L5 is "Weak (public information entirely blank)", and most other layers are mostly "Medium", with most ratings constrained by insufficient public information rather than confirmed capability;
• Summary positioning: Tencent Yuanqi is "AI Harness's channel / ecosystem encapsulation layer" — its value lies in distribution and monetization, not runtime and engineering.

2. Glossary

3. Feature Description

3.1. No-Code Agent Creation

Yuanqi's creation flow is a typical configuration-based (rather than orchestration-based) path:

1. Choose a model (Hunyuan family and third-party models are selectable);
2. Write a prompt and welcome message, optionally improved automatically with "One-click Optimize";
3. Attach a knowledge base, plugins, and workflows;
4. Test the conversational result;
5. Publish with one click to channels such as Official Account / QQ / Mini Program / Yuanbao / WeCom.

3.2. Knowledge Base System

Engineering detail tip (practical experience from third-party tutorials): when syncing authorized Official Account articles, one article per file works better than merging several articles. Official Account auto-sync is Yuanqi's unique customer-acquisition-side knowledge base capability — the knowledge base updates automatically as content updates, and this design precisely matches the real workflow of private-domain operation.

3.3. Plugins and Workflows

• Plugins: let an agent call external APIs, such as web search, document interpretation, and image generation; there is a Plugin Plaza entry;
• Workflows: visually combine plugins, large-model calls, and code blocks to implement more stable multi-step processes;
• Reference Display: replies show by default which knowledge base / plugin / workflow was invoked (can be disabled) — this is a rare, user-side transparency mechanism that can fall under L5 observation, but its depth goes no further than this.

3.4. WeChat Ecosystem Distribution and the Monetization Loop

This is Yuanqi's most structurally differentiated capability, broken into three parts:

1. Reach: agents can be published to the WeChat Official Account (including menu bar and in-article embedding), QQ, Mini Programs, WeCom, and Tencent Meeting;
2. Knowledge: authorized Official Account articles auto-sync into the knowledge base, so content grows in sync with the agent's capabilities;
3. Monetization: WeChat Pay MCP gives agents built-in payment capabilities (tipping, ordering), completing the "reach — serve — charge" loop.

From a Harness perspective, these are not three marketing features but rather the making of "distribution" into a platform-level capability — compared with Coze's one-click multi-channel publishing (see 06-coze.md), Yuanqi's channels are more concentrated within the Tencent ecosystem, but the payment loop is something Coze does not have.

3.5. Consumer-Grade Capabilities on the Yuanbao Side

• AI search (Official Account / Channels sources);
• Document parsing and Agent delivery (PPT / Word / Excel / PDF / HTML);
• Group chat summarization, deep thinking, code generation;
• Use directly by @Yuanbao within WeChat / QQ / Mini Programs;
• Multimodal: mutual translation across 33 languages including Chinese / English / Japanese, plus 5 Chinese minority languages such as Tibetan and Kazakh (third-party figures).

4. Platform Architecture

图 4-1｜腾讯元器 / 元宝平台可确认的三层架构

数据来源：基于本文分析绘制的示意图。

4.1. The Three-Layer Structure That Can Be Confirmed

This study must first declare: no official Tencent-published architecture documents, runtime design, state-management, or persistence mechanisms were found as public documentation. The following is only the form that can be confirmed from product behavior and third-party materials:

┌────────────────────────────────────────────────────────┐
│ 发布分发层（元器独有价值）                                 │
│  微信公众号 · QQ · 小程序 · 腾讯元宝 · 企业微信 · 腾讯会议  │
└────────────────────────────────────────────────────────┘
                          │
┌────────────────────────────────────────────────────────┐
│ 智能体配置层                                             │
│  智能体（提示词 + 欢迎语）· 知识库（三类）· 插件 · 工作流    │
│  参考源展示 · 智能体广场                                   │
└────────────────────────────────────────────────────────┘
                          │
┌────────────────────────────────────────────────────────┐
│ 模型层                                                   │
│  混元系列（版本口径冲突，见 4.2 节）+ 第三方模型              │
│  元宝侧：混元 Hy3 / Hy4 preview + DeepSeek 双引擎          │
└────────────────────────────────────────────────────────┘

4.2. Conflicting Claims at the Model Layer

The Hunyuan base versions given by different third-party sources are mutually incompatible; this study does not trim them but presents them in parallel:

The three claims may correspond to three different contexts — "platform base", "API billing tier", and "Yuanbao first release" — but because Tencent's official documentation is sparsely public, they cannot be further merged or disproven, marked [To be verified].

4.3. Processing Flow of an Official Account Agent Request

1. A fan asks a question inside the Official Account;
2. The agent loads its configuration: prompt, knowledge base (including synced Official Account articles), plugins, and workflows;
3. If there is a knowledge base hit: retrieve and inject into the context, and the reply includes a reference source display by default;
4. If plugins / workflows are configured: call external APIs or execute a multi-step flow;
5. If a transaction is involved: complete tipping / ordering via WeChat Pay MCP;
6. The reply is relayed back to the fan through the Official Account backend.

5. Harness Design

5.1. Six-Layer Capability Overview

Important note: this platform's six-layer ratings are constrained by "seriously insufficient public information". Apart from L5 (public information confirmed entirely blank after cross-checking multiple sources), the "Medium" on the other layers should be read as "evidence supports medium, no evidence supports higher", rather than "measured as medium".

5.2. L1 Context Engineering Layer

• The knowledge base is the main mechanism of this layer: three knowledge base types + Official Account article auto-sync cover most of the fact-source needs for private-domain operation scenarios;
• Prompt "One-click Optimize" is a lightweight context aid for non-technical users;
• Missing: no public context compression, retrieval-result priority ranking, or cache reuse mechanisms; token management in long-document scenarios is entirely invisible.

5.3. L2 Tool and Execution Layer

• The Plugin Plaza provides basic capabilities such as web search, document interpretation, and image generation;
• Workflow nodes can combine plugins, large-model calls, and code blocks;
• WeChat Pay MCP is the most valuable item in this layer: it turns "transactions" into a native tool of the agent, a capability unique among all 14 platforms in the group;
• Missing: no public information on sandbox execution or a code interpreter; whether plugins are based on standard MCP is also unspecified.

5.4. L3 Orchestration and Control Layer

• Visual workflows cover the "plugin + LLM + code block" combination, sufficient for customer service and Q&A-style flows;
• Third-party evaluations generally consider its complex business flow orchestration capability insufficient — no public documentation for sub-agent dispatch, human approval nodes, or loops and parallelism;
• No public information on a HITL (human-in-the-loop) mechanism.

5.5. L4 Memory and State Layer

• Third-party evaluations mention the platform has database / long-term memory capability;
• Official docs do not explain the specific mechanisms of multi-turn session management and long-term memory — whether memory exists, where it lives, and whether it can be retrieved across sessions are all unverifiable;
• No public information on artifact management or checkpoints.

5.6. L5 Evaluation and Observability Layer

This is the layer with the most blank public information among all 14 platforms, confirmed by cross-checking multiple sources:

• Present: only "Reference Display" — showing which knowledge base / plugin / workflow was invoked in replies; this is user-side transparency, not engineering-side observability;
• Completely absent: any public documents on evaluation sets, evaluators, Trace chains, regression evaluation, A/B testing, or quality metrics;
• Conclusion: developers cannot answer the basic question "is my agent performing well, and did my change make it worse" on the platform. Teams needing evaluation capability must fully bolt on external solutions, and even the precondition for that (that invocation data can be exported) has no public documentation.

5.7. L6 Governance and Security Layer

5.8. Concrete Manifestations of Three Inherent Tensions

6. Case Studies

No enterprise deployment case studies with client names and quantified effect data were found. As of the retrieval date 2026-09-12, public information consists mainly of building tutorials and platform capability descriptions from individuals / Official Account operators, with no quantified indicators such as cost reduction or efficiency gains; this is reported as-is and not filled in.

Reference platform-side signals (third-party agents showcased in the Agent Plaza):

• Personal blogger agents such as "吴川斌的博客AI助手" (Wu Chuanbin's Blog AI Assistant) and "卡尔的AI沃茨" (Karl's AI Watts);
• "熊猫花花" (Panda Huahua) (jointly created by the Xinhua News Agency client × the Global Panda Partners Conference × the Pandaful Panda Community × Tencent Hunyuan);
• Vertical-domain service agents such as "得理法律咨询" (Deli Legal Consultation).

These signals show the platform has real usage in Official Account scenarios such as media and legal consultation, but they do not constitute evidence of effectiveness.

Indirect evidence of the monetization loop: the integration of WeChat Pay MCP (2025-07) shows the platform side is driving agent monetization, but there is no public data on the actual transaction volume of tipping / ordering.

7. Summary

7.1. Strengths

1. Unique native integration with the WeChat / WeCom ecosystem: Official Account knowledge base auto-sync + one-click publishing + WeCom / Mini Program reach, which no competitor can replicate;
2. Complete monetization loop: WeChat Pay MCP lets agents tip and order directly, closing the "reach — serve — charge" loop in one go;
3. Extremely low no-code barrier: configuration-based creation + One-click Optimize prompts, one of the fastest launch paths for non-technical users;
4. Mature private-domain / customer service scenarios: Q&A-pair knowledge bases and Official Account article sync precisely match private-domain operation workflows;
5. Yuanbao completely free: no usage cost on the consumer entry point, with the official explicit statement that there are no current charging plans;
6. Strong Chinese-language capability of Hunyuan (vendor claim): the closeness of Chinese corpora and WeChat ecosystem data is its natural model-side advantage.

7.2. Weaknesses

1. L5 evaluation-and-observability public information is entirely blank: no evaluation sets, no Trace, no regression evaluation — the most blank layer among all 14 platforms in the group;
2. Extreme scarcity of technical documentation: architecture, state management, persistence, and memory mechanisms are all unpublished, so engineering teams cannot do reliable technical due diligence;
3. Complex orchestration capability considered insufficient: third-party evaluations generally report weak support for complex business flows;
4. Many conflicting claims: three coexisting Hunyuan version claims and two statements on private deployment support mean any parameter must be cited with caution;
5. Relatively closed ecosystem: value is highly bound to Tencent-system channels, and cross-ecosystem distribution capability is weak;
6. Questionable monetization sustainability: under the free-Yuanbao strategy, the platform-side monetization path is unclear (no public explanation of revenue beyond the Yuanqi enterprise version and API billing).

7.3. Applicability Boundaries

7.4. Selection Recommendations

• Positioning judgment: treat Yuanqi as a "channel / ecosystem encapsulation layer". When the business's core question is "how does the agent reach users in WeChat and complete transactions", it is the first choice; when the core question is "how is the agent engineered, evaluated, and governed", it is not the answer.
• Trade-off with Coze (both are channel-oriented Agent Platforms): for private-domain and transactions within the WeChat ecosystem, Yuanqi's Official Account sync + WeChat Pay loop is stronger; in the Feishu / Douyin ecosystem or when open-source private deployment is needed, Coze (see 06-coze.md) is more suitable.
• Combination strategy: use Yuanqi for WeChat-side reach and light service, put heavy reasoning and evaluation on other platforms, and connect them via API or workflows — but this depends on how open the platform's API is, which is not covered by the public information, and must be validated by testing before adoption.
• Must confirm before adoption: whether private deployment is truly available (the claim conflict needs clarification); whether invocation data can be exported to build self-run evaluation; and the official view on the Hunyuan version and billing tiers.

Information Gap Statement

1. L5 evaluation and observability capability: confirmed by cross-checking multiple sources to be entirely blank in public information — no official documents for evaluation sets, Trace, or regression evaluation. This is the platform's biggest information gap and the capability item most in need of official clarification.
2. Missing official technical architecture documents: runtime design, state management, and persistence mechanisms are all unpublished; the architecture diagram in Chapter 4 is only an induction of the confirmable form.
3. Exact launch date: only known to be 2024; the specific date was not found.
4. Conflicting claims on private deployment: one source says not supported, another says supported (extra cost).
5. Conflicting claims on Hunyuan versions: three claims are presented in parallel — Hunyuan 2.0 (406B / 32B), Hy3 (295B / 21B), and Hy4 preview (770B / 49B) — and could not be merged.
6. Enterprise case studies with quantified effects: not found, and not filled in.
7. Multi-turn session and long-term memory mechanisms: not explained officially; only third-party evaluations mention database / long-term memory capability.
8. Platform API openness: no public documentation on whether agent invocation data can be exported or deeply integrated via API.
9. Enterprise-pricing details: the third-party claim "usually CNY 50,000+/year" has not been officially confirmed.

8. References

1. Tencent Yuanqi official website. http://yuanqi.tencent.com
2. Create a dedicated Tencent Yuanqi AI agent for your WeChat Official Account with zero barrier — Tencent Cloud Developer Community. https://cloud.tencent.com/developer/article/2456683
3. Build an enterprise knowledge base Q&A assistant with Tencent Yuanqi — Tencent Cloud Developer Community. https://developer.cloud.tencent.com/article/2609061
4. Done in 5 minutes! The whole process of building my Official Account's exclusive database knowledge base — Tencent Cloud Developer Community. https://cloud.tencent.com/developer/article/2648182
5. Tencent Yuanqi — ChooseAI (Hunyuan 2.0 claims, pricing, WeChat Pay MCP). https://www.chooseai.net/tools/93/
6. What is Tencent Yuanbao? Features, pricing and comparison (2026) — All Agent encyclopedia (Hy3 / Hy4 claims, Yuanbao free-strategy verification). https://www.allagent.wiki/agents/tencent-yuanbao
7. 2026 enterprise agent development platform panoramic evaluation: horizontal comparison of eight mainstream platforms — Juejin (Class-3 MLPS, one private-deployment claim). https://juejin.cn/post/7654244323158016038
8. 2026 full-dimension evaluation of AI agent platforms: from "raising lobster" to enterprise-grade deployment — CSDN (third-party cost figures, second private-deployment claim). https://blog.csdn.net/weixin_56622231/article/details/159515126
9. Most complete, newest, most detailed panorama guide to Tencent AI products — CSDN. https://blog.csdn.net/weixin_49076592/article/details/163128796
10. Project Parameter Card v1.0 (Agent Platform definition and the six-layer capability model) — this project's internal baseline file.