China's State Telecoms Researcher on AI Agents: The Case, and His Own Caveats

Curated by: The Long View editors
muted documentary photography, diplomatic setting, formal atmosphere, institutional gravitas, desaturated color palette, press photography style, 35mm film grain, natural lighting, professional photojournalism, A bronze ceremonial meter embedded in a sandstone plinth, its face inscribed with official seals and Mandarin characters denoting "National AI Flow," thin streams of golden light pulsing like measured current through glass capillaries, side-lit by low-angle morning sun casting long institutional shadows, stillness in a vast, empty plaza at dawn, atmosphere of restrained authority and calibrated progress [fal-ai/z-image/turbo]
The label “first year of large-scale agent deployment” emerged in China’s industrial circles in early 2026, traced to a CAICT vice president’s lecture on CCTV, where agents were defined by four operational steps and a token economy scaled to 140 trillion daily uses—measured alongside acknowledged limits in reasoning, energy, and human oversight.
China's State Telecoms Researcher on AI Agents: The Case, and His Own Caveats In Plain English: During the World AI Conference in Shanghai in July 2026, Chinese state television aired a lecture on AI agents: software that does not just answer a question but carries out the task, such as booking a whole business trip across several apps. The lecturer, Wei Liang, is a vice president of the China Academy of Information and Communications Technology (CAICT), the research institute under China's Ministry of Industry and Information Technology that helps draft the country's technology standards and policy. His talk is a fair guide to how Beijing wants agents understood: as a new layer of economic infrastructure, measured in "tokens" the way electricity is measured in kilowatt-hours. He also spent a good part of the lecture on what agents cannot yet do. Summary: The lecture aired on 19 July 2026 on CCTV's China Economy Lecture Hall (中国经济大讲堂). The programme's narrator opens by saying that Chinese industry calls 2026 "the first year of large-scale agent deployment." That label comes from the programme, not from Wei, and he does not argue for it. What he does is describe what deployment looks like and what the state thinks should go with it. His definition is practical. Ordinary software makes you work through four apps by hand to arrange a trip: travel approval, train tickets, airport car, hotel. A chatbot writes you a plan. An agent, told "book my trip to Shanghai tomorrow within company hotel rules", submits the approval, books the tickets and the hotel, and files the expense claim when you get back. He breaks this into four steps that repeat: understand the request, plan, act through software tools, and review the result. His examples come from manufacturing, health and local government. A Shanghai company fitted a five-axis machine tool with an "AI brain" that reads a 3D model and plans its own machining; programming that took a skilled engineer a day now takes two hours, and he says aerospace and electric-vehicle makers use it for their most precise parts. In health, an agent first built in Zhejiang became Ant Group's Ant Afu (蚂蚁阿福), which he says links more than 5,000 hospitals and 300,000 doctors and has served over 100 million users, 84 million of them in third-tier cities and smaller places. In government, he cites a Shenzhen service assistant that walks a person through opening a restaurant while filling in the forms, and a Wuhan city-operations agent that spots and dispatches problems such as street obstructions and traffic jams. His larger claim is about economics. Agents, he argues, are creating a "token economy": because one instruction to an agent triggers many rounds of model calls, the token turns from a billing unit into the standard measure of the cost and output of AI work. "For the first time," he says, computing power becomes "like electricity", something that can be precisely metered and traded as a general input to production. His figures: China's daily token use rose from 100 billion in early 2024 to 140 trillion by the end of March 2026. On OpenRouter in June 2026, Chinese models processed 18.8 trillion tokens a week, more than three times the figure for US models. He adds that OpenRouter is one platform and not the whole market. Then he lists the problems, which is the most useful part for an outside reader. Agents hallucinate and lose the thread on long chains of reasoning, so medicine and finance still need human review. A complex agent task uses about 136 times the energy of a chatbot reply, which raises hardware and power costs. Checking an agent's output is expensive: an agent can draft a market report in ten minutes, but a person may then spend two to three hours checking sources and fixing logic, so the total can exceed the time to write it by hand. He calls this "false prosperity." He warns that some projects borrow the word "token" for speculation unrelated to real computing demand, and that compute supply is not keeping up. On safety he sets three rules for users: keep sensitive data on private or local systems, have people review important results, and give agents tiered permissions with every action logged for audit. He names three risks: agents misled by malicious instructions, untested open-source plugins and third-party models, and agents with broad system access becoming a route in for hackers. His advice to companies is to start with contract review, report generation or customer service and expand later. China, he says, is writing national and sector standards for agents and wants them adopted into international standards. Material from the programme, not Wei: The CCTV programme surrounds the lecture with its own reports. It says three government departments issued China's first dedicated policy on agents in May 2026, and cites an IDC forecast of 350 million active enterprise agents in China by 2031. It reports that a grid agent in Xiamen cut average repair times by 50 minutes during a July heatwave; that Huawei's Pangu industrial agent at JAC Motors covers more than 1,500 quality-inspection tasks with a 99.99% defect catch rate; and that China's enterprise agent market will reach 44.9 billion yuan in 2026, double the year before. It also quotes Xi Jinping at the conference on guiding AI with "human wisdom and international consensus." Why it matters: CAICT is where Chinese industrial policy for technology gets drafted, so a CAICT vice president explaining agents on CCTV during WAIC week is close to an official position. Two parts of it deserve attention in the West. The first is the framing of tokens as a metered production input, like power, which points to Beijing treating AI usage as something to measure, standardise and plan for. The second is the caveats. Verification cost, energy use and token speculation are the problems the state's own researchers are watching, and they are likely places for future rules. Key Points: - Wei Liang, a vice president of CAICT, presented the lecture on CCTV during the 2026 World AI Conference. - The "first year of large-scale agent deployment" label comes from the programme's narrator, attributed to industry, not from Wei. - Wei defines agents by four repeating steps: understand, plan, act, review. - He argues agents turn tokens into a metered input to production, comparable to electricity. - He lists limits: hallucination, weak long-chain reasoning, high energy use, and checking costs that can cancel out the time saved. - He sets three rules for users (data stays local, human review, tiered and logged permissions) and names three risks (misled agents, untested plugins, over-broad access). Notable Quotes: - "Agents are moving from 'help me think it through' to 'help me get it done.'" (Wei Liang) - Computing power becomes, "for the first time, like electricity", precisely metered and flexibly traded. (Wei Liang, paraphrased from the Mandarin) - Output that looks efficient can be "a false prosperity" if people then spend hours checking it. (Wei Liang, paraphrased) Data Points: - 100 billion to 140 trillion: China's daily token use, early 2024 to end of March 2026 (Wei). - 18.8 trillion: weekly tokens processed by Chinese models on OpenRouter in June 2026, over three times the US figure; one platform only (Wei). - 136 times: energy of a complex agent task compared with a chatbot reply, an industry estimate cited by Wei. - One day to two hours: machine-tool programming time at a Shanghai manufacturer (Wei). - Over 100 million: Ant Afu users, 84 million outside China's larger cities (Wei). - 350 million: IDC forecast of active enterprise agents in China by 2031 (programme). Controversial Claims: - Wei says the best coding agents already exceed most human programmers at software engineering. - Whether the "token economy" is a new economic form or a billing method with a new name. - The token-use figures come from Chinese and platform sources and have not been independently checked. Technical Terms: - AI agent (智能体): software that plans and carries out multi-step tasks using other software tools. - Token: the unit in which AI models process text and are billed. - Skill library (技能库): plug-in modules that give an agent abilities in a specific field, such as finance or medicine. - OpenRouter: a service that routes requests to many AI models and publishes usage figures. Source: CCTV, China Economy Lecture Hall, 19 July 2026, "Agents lead a new AI transformation" (智能体引领AI新变革). Mandarin; transcribed and translated by The Long View. https://www.youtube.com/watch?v=lQgGI-4BbLY
Published November 5, 2026