Recently, the World Artificial Intelligence Conference (WAIC) concluded, but the discussion on "where to go" for artificial intelligence has not reached its end. If 2025 is the year of reasoning models, 2026 is the year of AI agents. Domestic and international tech giants are conducting multi-pronged explorations and accelerating their land-grabbing around agents, while numerous small and medium-sized enterprises are also diving in to cultivate the field, firmly believing that agents will become the new blue ocean after the traditional internet.
At this year's WAIC, many companies revealed their development plans in the field of agents. Some are positioning themselves in infrastructure, some are anchoring on underlying technologies, and more companies are focusing on scenario implementation and building terminal products. From model competition to product competition, a new round of competition around AI agents has already begun.
Agents Are Everywhere
At previous conferences, the industry focus was more on model parameters and concept demonstrations. This year, however, agents have truly grown into the fabric of products. From mainstream consumer electronics like smartphones and PCs to cutting-edge terminals such as smart glasses, cars, and robots, and further to niche categories like stationery and toys, diverse hardware terminals have all made agent capabilities their core selling point. In terms of agent implementation, whether it is the breadth of product types, the depth of functional capabilities, or the maturity of commercialization, a qualitative leap has been achieved compared to previous years.
The smartphone is the screen carrier closest to the user and also the most natural Agent carrier. Hu Baishan, President of vivo Central Research Institute, stated that the transition from Smart Phone to Agent Phone is an inevitable industry trend—smartphones will transform from "passive tools" to "proactive partners," proactively providing decision support and even executing tasks based on continuous perception of the environment and users. He pointed out that this will be an important source of incremental growth for the smartphone market in the next three to five years. Fang Fei, President of Honor's Product Line, believes that the terminals of the next decade will no longer be "containers for applications," but "stages for agents."
This judgment is being corroborated by both policy and industry. Not long ago, seven models providing on-device generative AI services for smartphones, including Apple Intelligence, Huawei Xiaoyi AI large model, OPPO AndesGPT large model, vivo BlueLM on-device large model, Xiaomi HyperAI, Samsung Galaxy AI, and Nubia Doubao smartphone large model, passed the filing of the Cyberspace Administration of China. At WAIC, three products—StepFun STEPX Neo agent smartphone, Honor Robot Phone, and Nubia NaviX Ultra 2nd Gen Doubao smartphone—were unveiled on the same stage, with long queues of attendees eager to catch a glimpse. Manufacturers such as OPPO and vivo also announced that they will carry out systematic layouts around system-level agents.
Targeting more types of physical terminals, agent applications continue to extend across various scenarios. In the office scenario, Lenovo launched the Tianxi Personal Super Agent, featuring AI control, AI search, AI translation, AI notes, and AI services, which has now fully landed on its AI PC products. In the travel scenario, Eva, the "super agent" for cars jointly developed by Qianli Technology, Geely Auto, and StepFun, was officially unveiled on the mass-produced model Zeekr 8X, capable of completing complex operations such as natural interaction, itinerary planning, and local life service invocation. In the service scenario, 60 Agibot robots fully took over the exhibition operation and service work, undertaking functions such as guided tours, unmanned retail, interactive communication, and entertainment.
Agents at the software and service levels are also penetrating thousands of industries. Baidu's general-purpose agent "Baidu Dazi" can independently complete complex office tasks across applications and files; Yixin Group brought an Agentic large model for auto finance with scaled implementation; and Tec-Do released the Navos 2.0 marketing agent, aiming to solve many challenges in brand global marketing.
It can be seen that an obvious change is taking place: while models still determine the upper limit of an agent's capabilities, the true user-facing competition has shifted from the models themselves to agent products. The industry is no longer just answering "how strong the model is," but is further solving "how agents can truly serve users."
From "Terminal + AI" to "AI + Terminal"
After agents enter terminals, industrial competition is undergoing deeper changes. In the past few years, the industry has discussed "AI + smartphone," "AI + PC," and "AI + car" more frequently. The core of this model is to add AI capabilities to traditional terminals, which is essentially a functional superposition with the terminal as the body and AI as the application.
At this year's WAIC, more and more companies have begun to emphasize concepts such as "AI-native" and "Agentic OS." The so-called AI-native does not mean implanting a certain AI capability into a product, but defining the product around AI, and further defining the product around the Agent. As Meituan CEO Wang Xing said in the first-quarter earnings call, in the future, in addition to serving consumers (To C) and merchants (To B), serving AI Agents (To A) is becoming increasingly important.
The industry has also presented a clear dual-track layout: one is to delve inward, reconstructing the operating logic of the entire system from the underlying layer to meet the needs of agents; the other is to open up outward, promoting cross-vendor and cross-terminal ecological collaboration, allowing agents to break through the boundaries of a single device.
In terms of inward independent technology research and development, system-level agents have become the focus of attention. Honor announced the upgrade of MagicOS into a companion-type multimodal agent operating system—Agentic OS. Huang Fei, Honor's Chief AI Scientist, stated that the essence of Agentic OS is to reconstruct a new type of operating system centered on "intentions" and "tasks." The future Agentic OS will reconstruct the hardware layer, kernel layer, model layer, framework layer, interaction layer, and ecosystem layer as a whole. The objects of scheduling will change from processes and threads to intentions and tasks, and the resources managed by the kernel will expand from memory and computing power to models and agents. Alibaba Cloud, on the other hand, released the "Agent-Native Cloud Suite," specifically designed to manage, optimize, and orchestrate multi-agent workflows. Zhou Qi, head of Alibaba Cloud's Cloud Native Application Platform, believes: "The next round of competition is not about who has more Agents, but who can turn Agents into controllable, reusable, collaborative, and evolving organizational assets."
Opening up the external ecosystem allows the capability boundaries of agents to break free from the limitations of a single brand and a single device. The second-generation Doubao smartphone, jointly created by Nubia and ByteDance's Doubao, deeply injects the Doubao large model and agent scheduling capabilities into the system and hardware. StepFun announced its first batch of ecological cooperation partners, proposing to work with partners such as Alipay, Meituan, Amap, Didi, JD.com, and Baidu to build an agent ecosystem within 100 days. Various links, including model manufacturers, hardware terminals, and software manufacturers, are accelerating cross-border linkage, and an open and collaborative agent industrial ecosystem is gradually taking shape.
"Usable" Must Evolve into "Securely Usable"
Agents moving towards terminals also means that the AI industry is entering a new stage. Traditional AI assistants are more like information interaction tools, whereas agents possess higher execution privileges and may access more personal privacy and corporate information. Du Yuejin, Qiushi Distinguished Professor at Zhejiang University, pointed out that the industry has entered a stage of "problem explosion" in agent security. Agents integrate multiple links such as models, data, tools, permissions, and business processes. A vulnerability in any link will trigger chain risks. The single-model regulatory model can no longer adapt to the development of agents, and systematic regulation must be implemented. He emphasized that in critical fields such as finance, electric power, government affairs, and scientific research, a strict distinction must be made between "usable" and "securely usable."
From this perspective, establishing reliable technical standards and exploring trustworthy permission boundaries are not to restrict the development of agents, but are necessary prerequisites for them to move towards scaled applications.
Since this year, the construction of China's agent standard system has significantly accelerated. Policy deployment has gradually shifted from encouraging innovation to placing equal emphasis on innovation and regulation. On the one hand, it continues to encourage agent technology research and application development, supporting their implementation and flourishing in fields such as consumption, manufacturing, and services; on the other hand, it continuously improves institutional supply in terms of security governance, standard construction, and ecosystem building to safeguard production and living scenarios. Recently, relevant policies such as the "Basic Requirements for Security of Agent Applications" and the "Implementation Opinions on Regulated Application and Innovative Development of Agents," as well as a series of national standards such as "Intelligence Grading of AI Terminals" and "Interconnection of AI Agents," have been successively released.
Enterprise agreements and industry initiatives related to agents are also being released simultaneously, sending out positive signals for the standardized development of the industry. Ant Group, together with more than 20 industry partners including Zhipu AI, SenseTime, Xiaomi, and BYD, jointly promoted the construction of the Agent Security Link (ASL), a secure and trustworthy interconnection protocol for agents. Unlike protocols such as MCP and A2A that solve agent connection and invocation, ASL aims to establish a verifiable, transitive, constrainable, and auditable trust chain for cross-agent collaboration. It is reported that ASL has covered scenarios such as smartphone assistants, AI glasses, smart vehicle systems, and embodied intelligence.
Looking overseas, on July 27, NVIDIA announced the establishment of the Open Safe AI Coalition with more than 30 companies including Adobe, Hugging Face, Dell Technologies, Microsoft, SpaceX, and IBM. NVIDIA stated that it will open up models, model weights, datasets, and AI agent research achievements to help members track, audit, and manage AI agent behaviors.
Model capabilities determine the height of technology, security capabilities determine the breadth of applications, and the ecosystem model determines how far agents can ultimately go. Agents, which have moved from conceptual demos to the real world, possess infinite possibilities.
Author | Chen Cun Editor | Qiu Jiangyong Art Editor | Malia Supervisor | Wu Lilin