中国五年规划要求全经济领域部署 AI Agent：这意味着什么
英文标题：🔶 China's Five-Year Plan Mandates AI Agents Across the Economy: What It Means
中国第十五个五年规划（2026—2030年）拟投入超过1万亿元人民币发展 AI，其中约30%、约420亿美元将专门用于具备自主推理、决策和行动能力的智能体 AI。规划为制造、交通、医疗、服务和国防等领域设定了明确的 Agent 部署目标，并提出到2028年实现“4级以上自主性”。文章认为，这一国家级推动将形成大规模真实世界测试场，加速智能体能力、标准、算力和人才体系的发展。对开发者而言，模型多元化、开源可移植性以及安全护栏将成为跨地区部署 Agent 工作流的关键能力。
109.3 KB
中文
中国刚刚将 AI Agent 上升为国家层面的强制性部署方向。
将于2026年第一季度推出的第十五个五年规划（2026—2030年）承诺投入超过1万亿元人民币（约1400亿美元）发展 AI，其中约30%——约420亿美元——将专门用于智能体 AI（agentic AI）：即能够在极少人工监督下进行推理、决策和行动的自主系统。这不是一项研究愿景，而是一项带有目标、时间表和问责机制的产业政策。
对于全球 AI Agent 生态系统——包括 OpenClaw 等开源项目——这将改变竞争格局。
规划实际说了什么
国务院2026年2月发布的《人工智能高质量发展指南》为自主 Agent 的部署列出了具体目标：
**制造业：**到2028年，50%的指定“智能工厂”将由 Agent 运营。这意味着 AI 系统将自主负责生产线管理、质量控制、物流以及维护决策。
**交通运输：**到2027年建成全国性的 Robotaxi 网络，并在百度 Apollo 等自动驾驶平台基础上进一步发展。Agent 层不仅负责驾驶，还将处理车队协调、路径优化和乘客交互。
**医疗健康：**到2030年，80%的医院将采用 AI Agent 进行诊断和手术辅助。这类 Agent 系统不只是提出建议，而是在既定临床规范内采取行动。
**服务业：**目标是在2030年前部署超过1亿个家庭和办公 Agent。阿里巴巴的“Agent Ecosystem”计划及类似平台将成为主要交付机制。
**国防：**根据2026年国防白皮书，到2029年将把集群 Agent 整合进中国人民解放军的无人系统。
该规划的目标是在2028年前达到“4级以上自主性”——即 Agent 能够在动态环境中以接近人类独立性的水平运行。
企业机器已经开始运转
中国科技巨头并未等待相关指引生效：
- 百度截至2026年3月，已在10座城市部署 Ernie Agents，用于物流自动化
- 阿里巴巴的 Tongyi Agents 表示，其在电商工作流自动化中实现了40%的效率提升
- 字节跳动的 Doubao Swarm 为内容和推荐引擎运行多 Agent 系统
- 华为的 Celia Pro 提供企业级 Agent，并宣称其在工厂环境中的正常运行时间达到99%
国内 AI 模型生态也在同步成熟。DeepSeek-V3、Qwen-3 和华为 Pangu 模型提供了推理能力底座。国家互联网信息办公室于2026年1月出台的规定要求 Agent 必须“符合主权要求”——即基于部署在中国境内的数据和模型构建，关键基础设施不得依赖外国技术。
为什么这件事的影响超出中国
规模效应
当一个政府要求一个规模达18万亿美元经济体的所有主要行业采用 Agent 时，它就创造了全球最大的智能体 AI 真实世界试验场。数百万 Agent 在生产环境中运行，产生故障数据、边缘案例和优化洞察，所形成的反馈循环将使能力发展速度远超实验室环境所能实现的水平。
中国在2025年部署了超过60万个 Agent。按照五年规划的轨迹，到2028年这一数量将达到数百万。在这样的规模下，“能在演示中运行的 Agent”与“能在真实世界中运行的 Agent”之间的差距将迅速缩小。
标准之争
中国正通过自身渠道推动 Agent 互操作性标准。华为在 MWC 2026 上公布的 A2A-T 协议便是一个例子——这是一个开源 Agent 通信框架，旨在凭借庞大的采用规模成为事实标准。如果由中国开发的协议成为制造、物流和企业领域的标准，全球其他地区的 Agent 生态系统将需要与之实现互操作。
这与5G标准的策略如出一辙：先在国内实现大规模采用，再将标准输出至国际市场。
人才与算力投入
“东数西算”工程的扩展将为 Agent 训练分配百亿亿次浮点运算（exaFLOPS）级算力，目标是在边缘设备上运行10亿参数的 Agent 模型。结合本土芯片发展——包括受美国出口管制推动加速的壁仞 BR100 等芯片——中国正在构建一套从硅片到应用的垂直整合 Agent 技术栈。
世界知识产权组织（WIPO）数据显示，截至2025年，中国已拥有全球45%的 AI Agent 相关专利。五年规划的投资将进一步扩大这一领先优势。
这对 OpenClaw 用户意味着什么
模型多元化变得更加重要
AI Agent 的地缘政治维度真实存在。美国出口管制、Anthropic 获得五角大楼指定，以及中国的主权要求，都指向一个割裂的全球 AI 格局。OpenClaw 与模型无关的架构——可在 Claude、GPT、Gemini、DeepSeek、Qwen 和本地模型之间切换的能力——不只是一个便利功能，更是一种战略保险。
如果你正在构建可能需要跨司法辖区运行的 Agent 工作流，模型灵活性不是可选项。
开源是中立地带
OpenClaw 处于一个颇具意义的位置。作为可与任意模型提供商协作的开源框架，它是少数能够弥合中美 AI 分野的 Agent 平台之一。中国用户可以将其与国产模型搭配运行；西方用户则可以与西方提供商一起使用。Agent 逻辑——技能、记忆和工作流——可在两套生态系统之间迁移。
随着 AI 格局沿地缘政治边界进一步分裂，这种中立性可能会变得愈发有价值。
安全问题
中国的规划包含一项“AI 安全法”，要求高风险 Agent 必须保留人类参与决策（human-in-the-loop）机制。这与西方的担忧相呼应，但采取了不同的监管路径。对 Agent 构建者来说，结论具有普遍性：无论在哪个司法辖区，大规模自主 Agent 都需要安全护栏。
OpenClaw 的护栏配置、允许列表和权限系统与这一方向一致。随着全球 Agent 部署规模扩大，预计各地的安全要求都会收紧。
更大的图景
中国的五年规划是首个将 AI Agent 作为经济基础设施进行国家级押注的计划。美国拥有企业投资（OpenAI、Anthropic、Google）；欧盟拥有监管（AI Act）；中国则同时具备投资与带有明确目标的强制性采用政策。
结果是：2026—2030年将成为 AI Agent 从“有趣的工具”转变为“经济支柱”的十年——至少在一个主要经济体中如此。无论世界其他地区选择遵循强制推动模式还是市场驱动模式，方向已经确定。
对于 OpenClaw 的个人用户而言，实际含义很直接：**你正在基于一个刚被规模达18万亿美元的经济体宣布为关键基础设施的平台类别进行构建。**Agent 浪潮不是科技泡沫，而是产业转型。
问题不在于 AI Agent 是否会无处不在，而在于由谁构建、由谁控制，以及它们按谁的规则运行。对于自托管 AI的价值而言，这个问题如今更加紧迫。
English / 原文
China just made AI agents a national mandate.
The 15th Five-Year Plan (2026â2030), rolling out in Q1 2026, commits over Â¥1 trillion (~$140B USD) to AI development, with roughly 30% â about $42 billion â earmarked specifically for agentic AI : autonomous systems that reason, decide, and act with minimal human oversight. This isnât a research aspiration. Itâs industrial policy with targets, timelines, and accountability.
For the global AI agent ecosystem â including open-source projects like OpenClaw â this changes the landscape.
What the Plan Actually Says
The State Councilâs âGuidelines for High-Quality Development of Artificial Intelligenceâ (February 2026) lays out specific goals for autonomous agent deployment:
Manufacturing: 50% of designated âsmart factoriesâ to be agent-operated by 2028. This means AI systems managing production lines, quality control, logistics, and maintenance decisions autonomously.
Transportation: Nationwide robotaxi networks by 2027, building on Baidu Apollo and other autonomous vehicle platforms. The agent layer handles not just driving but fleet coordination, routing optimization, and passenger interaction.
Healthcare: 80% hospital adoption of AI agents for diagnostics and surgical assistance by 2030. Agent systems that donât just recommend but act within defined clinical protocols.
Services: A target of 100M+ deployed household and office agents by 2030. Alibabaâs âAgent Ecosystemâ initiative and similar platforms are the delivery mechanism.
Defense: Integration of swarm agents into PLA unmanned systems by 2029, per the 2026 defense white paper.
The plan targets âLevel 4+ autonomyâ by 2028 â meaning agents that operate with near-human independence in dynamic environments.
The Corporate Machine Is Already Moving
Chinaâs tech giants arenât waiting for guidelines to take effect:
- Baidu has deployed Ernie Agents across 10 cities for logistics automation as of March 2026
- Alibabaâs Tongyi Agents report 40% efficiency gains in e-commerce workflow automation
- ByteDanceâs Doubao Swarm runs multi-agent systems for content and recommendation engines
- Huaweiâs Celia Pro delivers enterprise agents with 99% uptime claims in factory settings
The domestic AI model ecosystem is maturing in parallel. DeepSeek-V3, Qwen-3, and Huaweiâs Pangu models provide the reasoning backbone. The Cyberspace Administrationâs January 2026 regulations require agents to be âsovereign-compliantâ â built on China-hosted data and models, with foreign tech dependencies banned for critical infrastructure.
Why This Matters Beyond China
The Scale Effect
When a government mandates agent adoption across every major industry in a $18 trillion economy, it creates the largest real-world testing ground for agentic AI. The feedback loops â millions of agents operating in production, generating failure data, edge cases, and optimization insights â will accelerate capability development far beyond what lab environments can produce.
China deployed 600,000+ agents in 2025. The Five-Year Plan trajectory suggests millions by 2028. At that scale, the gap between âagents that work in demosâ and âagents that work in the real worldâ closes fast.
The Standards Race
China is pushing agent interoperability standards through its own channels. Huaweiâs A2A-T protocol, announced at MWC 2026, is one example â an open-source agent communication framework designed to become a de facto standard through sheer adoption volume. If Chinese-developed protocols become the standard in manufacturing, logistics, and enterprise, the rest of the worldâs agent ecosystem will need to interoperate.
This echoes the 5G standards playbook: establish domestic adoption at scale, then export the standard internationally.
The Talent and Compute Investment
The âEast Data, West Computeâ program extension allocates exaFLOPS-scale compute specifically for agent training, with a goal of running 1-billion-parameter agent models on edge devices. Combined with domestic chip development (Biren BR100 and others accelerated by US export controls), China is building a vertically integrated agent stack â from silicon to application.
WIPO data shows China already holds 45% of global AI agent-related patents as of 2025. The Five-Year Plan investment will widen that lead.
What This Means for OpenClaw Users
Model Diversification Gets More Important
The geopolitical dimension of AI agents is real. US export controls, Anthropicâs Pentagon designation, and Chinaâs sovereignty requirements all point to a fragmented global AI landscape. OpenClawâs model-agnostic architecture â the ability to switch between Claude, GPT, Gemini, DeepSeek, Qwen, and local models â isnât just a convenience feature. Itâs strategic insurance.
If youâre building agent workflows that might need to operate across jurisdictions, model flexibility isnât optional.
Open-Source as Neutral Ground
OpenClaw sits in an interesting position. As an open-source framework that works with any model provider, itâs one of the few agent platforms that can bridge the US-China AI divide. Chinese users can run it with domestic models; Western users can run it with Western providers. The agent logic â skills, memory, workflows â is portable across both ecosystems.
This neutrality may become increasingly valuable as the AI landscape fragments along geopolitical lines.
The Safety Question
Chinaâs plan includes an âAI Safety Lawâ mandating human-in-the-loop for high-risk agents. This parallels Western concerns but takes a different regulatory approach. For agent builders, the takeaway is universal: autonomous agents at scale require safety guardrails, regardless of jurisdiction.
OpenClawâs guardrails configuration, allowlists, and permission systems align with this direction. As agent deployment scales globally, expect safety requirements to tighten everywhere.
The Bigger Picture
Chinaâs Five-Year Plan is the first national-scale bet on AI agents as economic infrastructure. The US has corporate investment (OpenAI, Anthropic, Google); the EU has regulation (AI Act); China has both investment and mandated adoption with specific targets.
The result: 2026â2030 will be the decade where AI agents move from âinteresting toolâ to âeconomic backboneâ â at least in one major economy. Whether the rest of the world follows the mandate model or the market model, the direction is set.
For individual OpenClaw users, the practical implication is straightforward: youâre building on a platform category that a $18 trillion economy just declared essential infrastructure. The agent wave isnât a tech bubble â itâs industrial transformation.
The question isnât whether AI agents will be everywhere. Itâs who builds them, who controls them, and on whose terms they operate. For the case for self-hosted AI, that question is even more urgent now.