In May 2026, a clear watershed moment emerged in the domestic AI agent track: previously, everyone was competing over "how large the model parameters are" and "how low the token cost has dropped," but now the competition has shifted to "how many agents are actually running in enterprises." This shift, more than any funding news, better reflects the maturity of the industry.
1. ByteDance: Turning AI Agents into a "Full-Modal Package"
On May 11, ByteDance released the industry's first full-modal intelligent agent package, Agent Plan. The significance of this move lies in: no longer selling model capabilities individually, but packaging voice, image, text, and video understanding into a "toolbox that enterprises can use on demand."
200 billion AI infrastructure capital expenditure in 2026 (upwardly revised from 160 billion) Release date of the enterprise-level AI office assistant "ByteDance Smart Assistant" on May 8
Another detail worth noting: ByteDance released its enterprise-level AI office assistant "ByteDance Assistant" on May 8, directly competing with Microsoft Copilot. The timing was very precise—just one day before the Microsoft Build conference, clearly aiming to capture attention in the enterprise market.
But ByteDance's problem is also obvious: its ToB gene is not as deep as BAT's, and enterprise customers' trust in "Made by ByteDance" still lags behind that of "Alibaba Cloud" and "Tencent Cloud". Whether the Agent Plan package can truly open up the enterprise market will be first tested by the data in Q3 2026.
2. Baidu: Changed a unit of measurement, shifting from Token to DAA
Baidu proposed a new concept in May: DAA (Daily Active Agents), to replace the traditional token consumption metric. This move is clever—Token is a technical metric that CIOs cannot understand; DAA is a business metric that CIOs can directly report to CEOs.
"To measure the prosperity of a platform, the core is to see how many Agents are working for humans and delivering results." — Robin Li, May 2026
Baidu has simultaneously upgraded its intelligent cloud to a "new full-stack AI cloud for large-scale intelligent agent applications," with the approach being a four-layer full stack of "chip, cloud, model, and entity." This concept is very similar to Microsoft's Azure AI stack, with the difference being that Baidu has an additional "chip" layer (the large-scale validation of the Kunlun Core P800).
Baidu Miaoda: 90% code auto-generation rate
Baidu's code agent "Miaoda" showed a code auto-generation rate of 90% in public data from May. If this figure is accurate, it means that for a moderately complex enterprise application, what previously required 10 developers can now be accomplished by 1 product manager and 1 AI agent.
However, the metric of "code auto-generation rate" itself needs to be viewed with caution—whether the generated code can run directly or requires significant manual revision is the true measure of value. Baidu has not yet released more detailed breakdown data on this part.
3. Alibaba: Tongyi Qianwen + Full E-commerce Chain, Commercialization Enters Return Period
On May 11, Alibaba fully integrated Tongyi Qianwen with Taobao and Tmall. This move is not a technology showcase but a delivery of commercial value—when a user asks on Tmall, "Is this suitable for me?", the AI does not just spit out a block of text; instead, it directly retrieves historical purchase data, body measurement data, and review data to provide a "suitable/unsuitable" judgment.
40% improvement in shopping decision efficiency (Alibaba official data) 90%+ Wukong Agent platform renewal rate
A more important signal is that the renewal rate of the "Wukong" Agent platform has exceeded 90%. The renewal rate, as an indicator, speaks more to product value than any financing news—companies are willing to renew, which shows that the intelligent agent is truly generating business value, rather than just being a novelty in the POC (Proof of Concept) stage.
Alibaba's advantage lies in the fact that e-commerce scenarios are one of the areas where the value of AI agents is easiest to quantify. An agent that saves a user 10 minutes of selection time is a value that enterprise customers can directly perceive. In contrast, quantifying the value of ToB ERP agents is much more difficult.
4. Tencent: WorkBuddy becomes the most widely used productivity agent in China
Tencent did not release new products as intensively as several others in May, but one figure remained steady: WorkBuddy is the most widely used productivity agent service in China (by number of enterprise customers).
In his internal speech in May, Ma Huateng mentioned that "AI's efforts to shore up weaknesses have shown results" — behind this statement lies a clear shift in Tencent's AI strategy: no longer pursuing leadership in model parameters, but instead deeply embedding AI capabilities into billion-user products like WeChat, WeCom, and Tencent Meeting.
This approach is very similar to Microsoft's: Microsoft does not pursue exclusivity with GPT-5, but instead embeds Copilot into Word, Excel, and Outlook. Tencent's logic is the same: it does not pursue a large model brand like Tongyi Qianwen, but aims to have hundreds of millions of users using AI features every day (even if they may not feel "this is AI").
5. 360 and Huawei: Differentiated Approaches
360 Security + Inclusiveness
Lobster Plan · Lightweight Version Saves 99% Tokens
- Launch the "Lobster Plan" and distribute 100 million Tokens to each employee
- Launch a lightweight version, saving up to 99% of token consumption
- Publish OpenClaw Ecological Security Risk Analysis
- Upgrade to AI Agent Ecosystem Automated Security Audit
Huawei Industry + Terminal
Agent Factory · CodeDao commercially launched on May 30
- Agent Factory solution released on May 8
- Xiaoyi Claw upgrade, based on HarmonyOS 6.0
- CodeDao Code Agent will be commercially available on May 30
- E-commerce development cycle shortened from 6 weeks to 3 weeks
360's strategy is "security + inclusiveness": security is 360's traditional advantage, while inclusiveness means driving token costs extremely low (if the 99% savings figure is true, it would be industry-shaking). Huawei's approach is "industrial scenarios": its intelligent agent factory solution directly targets the manufacturing sector, a track where BAT currently has relatively shallow coverage.
6. Horizontal Comparison: Where Does Each Company Excel?
| Manufacturer | Core Positioning | Key Actions in May | Advantage Track | Pending breakthrough |
|---|---|---|---|---|
| ByteDance | Full-Modal Package Bundle | Agent Plan + ByteSmart | Office collaboration, content generation | Enterprise Customer Trust |
| Baidu | Full-stack AI Cloud (Chip-Cloud-Model-Body) | DAA new measurement system + Second Da 90% code generation | AI-native application development | Renewal rate data not disclosed |
| Alibaba | E-commerce + enterprise services dual-wheel | Tongyi × Taobao Tmall + Wukong 90% renewal | E-commerce, Enterprise Agent Platform | Internationalization progress is slow |
| Tencent | The invisible AI embedded in billion-level products | WorkBuddy's most extensive + Ma Huateng's internal speech | Collaborative office, social | Large model brand voice |
| 360 | Safety + Inclusiveness | Lobster Plan + Save 99% Tokens | Security audit, small and medium-sized enterprises | Standardized Product Maturity |
| Huawei | Industrial scenario differentiation | Agent Factory + CodeDao Commercial Launch on May 30 | Industrial manufacturing, terminals | Ecological openness |
7. A Question Worth Pondering
After reviewing the May updates from six vendors, one question emerges: is the gap between domestic AI agents and their foreign counterparts (Microsoft/Google/SAP) narrowing or widening?
From the perspective of model capabilities, the gap is narrowing. Baidu's ERNIE 5.1, Alibaba's Tongyi Qianwen, and Tencent's models are already on par with Claude and GPT in Chinese-language scenarios.
But from the perspective of the "agent governance system," the gap may be widening. Microsoft Agent 365 went GA on May 1, followed closely by Google's AI Control Center — they are already addressing the issue of "how to manage 100 agents." Among domestic manufacturers, including the top six, the weight given to "governance" in publicly released materials is far lower than that given to "capability."
This gap will not affect order acquisition in the short term, but it will determine customer renewal rates and reputation 12 to 18 months later. For vendors with strong governance, the number of customer agents will naturally grow (because CIOs will approve more scenarios). For vendors with poor governance, the entire project may be halted as soon as the first security incident occurs.
CIOs who experienced the ERP era should remember: from 2005 to 2010, ERP vendors competed on "whether the features were complete"; from 2010 to 2015, they competed on "whether it was stable and compliant." AI agents may be following a similar path—only the cycle will be compressed to within 24 months.
8. Conclusion: What to Watch After May?
After the round of manufacturer updates in May 2026, there are actually only two indicators worth tracking:
- DAA (Daily Active Agents): This metric proposed by Baidu, if also disclosed by other manufacturers, can horizontally compare the real usage activity of various platforms. It's useless to just look at "how many features have been released"; instead, we need to see "how many agents are running every day."
- Renewal rate: The 90%+ announced by Alibaba Wukong is a good reference point. If other vendors also announce their renewal rates in Q3, it will reveal who is delivering value and who is just packaging concepts.
