 |
| Pix credit here |
The Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents, 智能体规范应用与创新发展实施意见, is a Chinese policy document formulated to implement the State Council's Opinions on Deeply Implementing the "Artificial Intelligence Plus" Initiative. It addresses "intelligent agents"—defined as intelligent systems possessing autonomous perception, memory, decision-making, interaction, and execution capabilities—as a distinct and important form of artificial intelligence products and services. The document is structured in six parts: fundamental principles, consolidating the development foundation, safeguarding the security baseline, strengthening application-led advancement, building an innovation ecosystem, and safeguard measures.
Read through the framework of New Era Modernization, the Implementation Opinions may be better understood not as an isolated technology-regulation document but as the latest articulation of the modernizing project at a specific point in its recursive unfolding. Chinese style modernization functions as the pre-given horizon of meaning — the Lebenswelt in Husserl's technical sense — within which Chinese constitutionalism constitutes itself and produces meaning, and that the CPC's periodic reformulation of the principal contradiction is the mechanism by which this horizon recalibrates without being abandoned. The Implementation Opinions sit within this framework as an instrument through which the party-state extends the modernizing project into a new technological domain — intelligent agents — while preserving the structural grammar of development-security pairing, Party-directed governance, and the juridification cascade from constitutional mandate through legislation to administrative practice.
Fundamental Principles
The document articulates four foundational principles: (1) safety and controllability, treating agent safety, reliability, and trustworthiness as baseline requirements throughout the entire technology lifecycle; (2) orderly regulation, constructing a governance system that connects smoothly with existing policies while establishing clear baseline and red-line boundaries; (3) innovation-driven development, pursuing systematic breakthroughs in key technologies through government-industry-academia collaboration; and (4) application-led advancement, leveraging representative use cases to drive technology verification and product iteration in a step-by-step manner.
These four principles are not merely policy preferences. They enact the development-security dialectic that can be understood as "a core feature of New Era governance: modernization operates as both engine and vessel for the rationalization of the state." Safety-and-controllability and orderly regulation pair structurally with innovation-driven development and application-led advancement — the document does not treat safety as a constraint on innovation or innovation as a risk to be managed; rather, both are presented as mutually constituting elements of the same governing project. Within the Marxist-Leninist framework that traces from Lenin's electrification thesis through Chinese modernization, this pairing reflects the vanguard party's claimed capacity to direct technological transformation as a conscious, planned process rather than an organic evolution — a project to be organized, implemented, and managed by a centralized authority that bends historical development to a conscious human purpose. The concept of "application-led advancement" (应用牵引) — literally "application traction" or "pull" — encodes a specific theory of innovation in which deployment needs drive technological development, consistent with the broader New Era emphasis on new quality productive forces (新质生产力).Technology and Standards
The Implementation Opinions mandate the strengthening of foundational technology research, including general-purpose models, specialized industry models, high-quality datasets, and key agent capabilities such as task comprehension, task planning, tool usage, long-term memory, and swarm collaboration. They call for a complete intelligent agent toolchain encompassing perception, memory, decision-making, interaction, and execution components, as well as security and governance tools such as adversarial sample detection and behavioral anomaly detection. An intelligent agent standards system is to be established, covering key technologies, data exchange, application scenarios, quality evaluation, security assurance, and trustworthy certification, with mandatory standards supported in healthcare, transportation, media, and public safety. Notably, the document envisions an "intelligent interconnection network" featuring agent registration platforms, digital identity management, search and discovery services, and interoperability protocols leveraging IPv6.
Security and Governance Framework
Section III establishes a multi-layered security architecture. Product standards require that intelligent agent behavior conform to laws, regulations, and mainstream values, with specific protections against algorithmic exploitation, dissemination of unwholesome values, and risks of addiction and emotional dependency among minors and the elderly. Decision-making authority must be clearly delineated between user-only decisions, user-authorized decisions, and agent-autonomous decisions, with users retaining the right to be informed and to exercise final authority. Technologies such as rule embedding and behavioral guardrails are mandated, and blockchain-based mechanisms are envisioned for verifiable and traceable agent behavior in important scenarios.
On security risk prevention, the document addresses intrinsic security capabilities (data security, cryptographic defense, attack detection, permission management), supply chain security across the full development lifecycle, and application-derived risks including automated attacks, privacy violations, false information, and cyber fraud. A "classified and tiered governance framework" differentiates between sensitive domains (subject to registration, testing, and defective product recall under joint oversight by the cyberspace administration and sectoral regulators) and low-risk domains (governed through compliance self-testing, information reporting, and industry self-discipline). Credit evaluation mechanisms for market participants are proposed, covering technology misuse, consumer inducement, false advertising, and concealment of defect information.
Application Domains
The document envisions intelligent agent deployment across an extraordinarily broad range of sectors: scientific research and R&D assistance; intelligent manufacturing, energy, transportation, agriculture, and financial services; consumer-facing applications in terminal devices, culture and tourism, and commercial services; education, healthcare, human resources, and information services; and government services, judicial services, public safety, urban governance, and bidding and tendering. This scope reflects the treatment of intelligent agents not as a narrow product category but as a cross-cutting technology embedded in the state's modernization project.
Innovation Ecosystem
To support deployment, the document promotes open-source innovation through domestic AI communities, compatibility with open-source chips and operating systems, and participation in open-source frameworks and toolchains. Industrial collaboration platforms, intelligent agent app stores, supply-and-demand matching activities, and demonstration projects in industrial clusters are envisioned. The document also contemplates cultivating a global ecosystem through international platforms such as the World Artificial Intelligence Conference, including overseas compliance frameworks adapted to local laws and cultural customs.
Analytical Insights
Analysis appears to yield several key insights:
AI as Strategic Infrastructure. The document's cross-sectoral scope suggests a characterization of China as treating AI "as strategic infrastructure for socialist modernization," in contrast to the EU's rights-and-risk framing or the U.S.'s fragmented sectoral approach.
The Development-Security Dialectic. The document pairs innovation-driven development with safety-and-controllability not as competing goals in a liberal balancing sense, but as "mutually constituting elements of the same governing project," where "development is pursued through mechanisms that embed security, and security is designed to enable sustainable development under state supervision".
Platforms as Governance Intermediaries. Development and distribution platforms are assigned the role of operationalizing party-state directives—establishing platform rules, user service agreements, and privacy policies and participating in credit evaluation—consistent with the observation that large technology companies function as "delegated governance" intermediaries rather than mere self-regulators.
Classified Governance as Administrative Allocation. The document's tiered framework is organized around administrative authority, not abstract risk levels, distinguishing it from the EU AI Act's risk-based classification, and instead reflecting what might be described as an allocation of administrative responsibility and inspection intensity.
The Instrumentalization Contradiction. The tension between deploying intelligent agents as governance tools (in government approvals, judicial case-handling, and public safety) and insisting on human controllability over those same agents reflects what I have elsewhere identified as the fundamental issue: "the more autonomous the tech, the greater the risk that the relationship between instrument and its wielders will be reversed".
Regulatory Layering. The Implementation Opinions add a new layer to China's existing regulatory ecology—including the Cybersecurity Law, Data Security Law, Personal Information Protection Law, algorithmic recommendation rules, deep synthesis provisions, and generative AI interim measures—but do not specify how agent-specific regulation will interact with these existing instruments.
Open Questions
The analysis identifies several unresolved issues: how the Implementation Opinions will coordinate with overlapping supervisory frameworks, including the MPS Measures on Cyberspace Security Supervision (Order No. 176); whether the intelligent interconnection network will evolve from a conventional infrastructure-standardization project into the kind of "operating system" for theorized normative orders; whether the "voluntary" credit evaluation mechanism will assume compulsory characteristics aligned with broader social credit systems; and whether the security imperatives will, in practice, predominate over development goals despite the document's dialectical framing.
The translation of ans analysis follows below.
Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents
智能体规范应用与创新发展实施意见
Original Text, Translation, and Analysis
With an Analysis Through the Prior Writing of Larry Catá Backer on Chinese Tech and AI Governance
ABSTRACT: This paper presents the original Chinese text, an English translation, and a scholarly analysis of the Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents (智能体规范应用与创新发展实施意见), a policy document issued pursuant to the State Council's "Artificial Intelligence Plus" initiative. The Implementation Opinions establish a governance framework for intelligent agents—defined as autonomous systems with capacities for perception, memory, decision-making, interaction, and execution—across scientific research, industrial production, energy, transportation, agriculture, financial services, education, healthcare, government approvals, judicial case-handling, public safety, and urban governance. The analytical portion identifies several key themes: the treatment of AI as strategic infrastructure for socialist modernization rather than merely a product-safety concern; a "development-security dialectic" in which innovation and control are presented as mutually constituting rather than competing imperatives; the assignment of governance-intermediary roles to development and distribution platforms; classified and tiered governance organized around administrative authority rather than abstract risk levels; and the instrumentalization contradiction—the tension between deploying autonomous agents as governance tools and maintaining party-state control over those same tools. The analysis also situates the Implementation Opinions within China's existing layered regulatory ecology, including the Cybersecurity Law, Data Security Law, algorithmic recommendation rules, deep synthesis provisions, and generative AI interim measures, and identifies unresolved questions of regulatory coordination and interaction.
EXECUTIVE SUMMARY:
The Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents, 智能体规范应用与创新发展实施意见, is a Chinese policy document formulated to implement the State Council's Opinions on Deeply Implementing the "Artificial Intelligence Plus" Initiative. It addresses "intelligent agents"—defined as intelligent systems possessing autonomous perception, memory, decision-making, interaction, and execution capabilities—as a distinct and important form of artificial intelligence products and services. The document is structured in six parts: fundamental principles, consolidating the development foundation, safeguarding the security baseline, strengthening application-led advancement, building an innovation ecosystem, and safeguard measures.
Read through the framework of New Era Modernization, the Implementation Opinions may be better understood not as an isolated technology-regulation document but as the latest articulation of the modernizing project at a specific point in its recursive unfolding. Chinese style modernization functions as the pre-given horizon of meaning — the Lebenswelt in Husserl's technical sense — within which Chinese constitutionalism constitutes itself and produces meaning, and that the CPC's periodic reformulation of the principal contradiction is the mechanism by which this horizon recalibrates without being abandoned. The Implementation Opinions sit within this framework as an instrument through which the party-state extends the modernizing project into a new technological domain — intelligent agents — while preserving the structural grammar of development-security pairing, Party-directed governance, and the juridification cascade from constitutional mandate through legislation to administrative practice.
Fundamental Principles
The document articulates four foundational principles: (1) safety and controllability, treating agent safety, reliability, and trustworthiness as baseline requirements throughout the entire technology lifecycle; (2) orderly regulation, constructing a governance system that connects smoothly with existing policies while establishing clear baseline and red-line boundaries; (3) innovation-driven development, pursuing systematic breakthroughs in key technologies through government-industry-academia collaboration; and (4) application-led advancement, leveraging representative use cases to drive technology verification and product iteration in a step-by-step manner.
These four principles are not merely policy preferences. They enact the development-security dialectic that can be understood as "a core feature of New Era governance: modernization operates as both engine and vessel for the rationalization of the state." Safety-and-controllability and orderly regulation pair structurally with innovation-driven development and application-led advancement — the document does not treat safety as a constraint on innovation or innovation as a risk to be managed; rather, both are presented as mutually constituting elements of the same governing project. Within the Marxist-Leninist framework that traces from Lenin's electrification thesis through Chinese modernization, this pairing reflects the vanguard party's claimed capacity to direct technological transformation as a conscious, planned process rather than an organic evolution — a project to be organized, implemented, and managed by a centralized authority that bends historical development to a conscious human purpose. The concept of "application-led advancement" (应用牵引) — literally "application traction" or "pull" — encodes a specific theory of innovation in which deployment needs drive technological development, consistent with the broader New Era emphasis on new quality productive forces (新质生产力).
Technology and Standards
The Implementation Opinions mandate the strengthening of foundational technology research, including general-purpose models, specialized industry models, high-quality datasets, and key agent capabilities such as task comprehension, task planning, tool usage, long-term memory, and swarm collaboration. They call for a complete intelligent agent toolchain encompassing perception, memory, decision-making, interaction, and execution components, as well as security and governance tools such as adversarial sample detection and behavioral anomaly detection. An intelligent agent standards system is to be established, covering key technologies, data exchange, application scenarios, quality evaluation, security assurance, and trustworthy certification, with mandatory standards supported in healthcare, transportation, media, and public safety. Notably, the document envisions an "intelligent interconnection network" featuring agent registration platforms, digital identity management, search and discovery services, and interoperability protocols leveraging IPv6.
Security and Governance Framework
Section III establishes a multi-layered security architecture. Product standards require that intelligent agent behavior conform to laws, regulations, and mainstream values, with specific protections against algorithmic exploitation, dissemination of unwholesome values, and risks of addiction and emotional dependency among minors and the elderly. Decision-making authority must be clearly delineated between user-only decisions, user-authorized decisions, and agent-autonomous decisions, with users retaining the right to be informed and to exercise final authority. Technologies such as rule embedding and behavioral guardrails are mandated, and blockchain-based mechanisms are envisioned for verifiable and traceable agent behavior in important scenarios.
On security risk prevention, the document addresses intrinsic security capabilities (data security, cryptographic defense, attack detection, permission management), supply chain security across the full development lifecycle, and application-derived risks including automated attacks, privacy violations, false information, and cyber fraud. A "classified and tiered governance framework" differentiates between sensitive domains (subject to registration, testing, and defective product recall under joint oversight by the cyberspace administration and sectoral regulators) and low-risk domains (governed through compliance self-testing, information reporting, and industry self-discipline). Credit evaluation mechanisms for market participants are proposed, covering technology misuse, consumer inducement, false advertising, and concealment of defect information.
Application Domains
The document envisions intelligent agent deployment across an extraordinarily broad range of sectors: scientific research and R&D assistance; intelligent manufacturing, energy, transportation, agriculture, and financial services; consumer-facing applications in terminal devices, culture and tourism, and commercial services; education, healthcare, human resources, and information services; and government services, judicial services, public safety, urban governance, and bidding and tendering. This scope reflects the treatment of intelligent agents not as a narrow product category but as a cross-cutting technology embedded in the state's modernization project.
Innovation Ecosystem
To support deployment, the document promotes open-source innovation through domestic AI communities, compatibility with open-source chips and operating systems, and participation in open-source frameworks and toolchains. Industrial collaboration platforms, intelligent agent app stores, supply-and-demand matching activities, and demonstration projects in industrial clusters are envisioned. The document also contemplates cultivating a global ecosystem through international platforms such as the World Artificial Intelligence Conference, including overseas compliance frameworks adapted to local laws and cultural customs.
Analytical Insights
Analysis appears to yield several key insights:
AI as Strategic Infrastructure. The document's cross-sectoral scope suggests a characterization of China as treating AI "as strategic infrastructure for socialist modernization," in contrast to the EU's rights-and-risk framing or the U.S.'s fragmented sectoral approach.
The Development-Security Dialectic. The document pairs innovation-driven development with safety-and-controllability not as competing goals in a liberal balancing sense, but as "mutually constituting elements of the same governing project," where "development is pursued through mechanisms that embed security, and security is designed to enable sustainable development under state supervision".
Platforms as Governance Intermediaries. Development and distribution platforms are assigned the role of operationalizing party-state directives—establishing platform rules, user service agreements, and privacy policies and participating in credit evaluation—consistent with the observation that large technology companies function as "delegated governance" intermediaries rather than mere self-regulators.
Classified Governance as Administrative Allocation. The document's tiered framework is organized around administrative authority, not abstract risk levels, distinguishing it from the EU AI Act's risk-based classification, and instead reflecting what might be described as an allocation of administrative responsibility and inspection intensity.
The Instrumentalization Contradiction. The tension between deploying intelligent agents as governance tools (in government approvals, judicial case-handling, and public safety) and insisting on human controllability over those same agents reflects what I have elsewhere identified as the fundamental issue: "the more autonomous the tech, the greater the risk that the relationship between instrument and its wielders will be reversed".
Regulatory Layering. The Implementation Opinions add a new layer to China's existing regulatory ecology—including the Cybersecurity Law, Data Security Law, Personal Information Protection Law, algorithmic recommendation rules, deep synthesis provisions, and generative AI interim measures—but do not specify how agent-specific regulation will interact with these existing instruments.
Open Questions
The analysis identifies several unresolved issues: how the Implementation Opinions will coordinate with overlapping supervisory frameworks, including the MPS Measures on Cyberspace Security Supervision (Order No. 176); whether the intelligent interconnection network will evolve from a conventional infrastructure-standardization project into the kind of "operating system" for theorized normative orders; whether the "voluntary" credit evaluation mechanism will assume compulsory characteristics aligned with broader social credit systems; and whether the security imperatives will, in practice, predominate over development goals despite the document's dialectical framing.
Contents
Part I: Original Chinese Text
Part II: English Translation
Part III: Analysis
PART I: ORIGINAL CHINESE TEXT
智能体规范应用与创新发展实施意见
智能体是具备自主感知、记忆、决策、交互与执行能力的智能系统，是人工智能产品及服务的重要形态。随着大模型等新一代人工智能技术迅猛发展，智能体正加速与网络空间、物理世界深度融合，深刻改变人类生产生活方式和社会治理模式。为落实国务院《关于深入实施“人工智能+”行动的意见》，促进智能体规范应用与创新发展，制定本实施意见。
一、基本原则
以推动科技创新、提升治理能力、构建产业生态、增进民生福祉为导向，坚持安全可控，将智能体安全、可靠、可信作为发展的底线要求，贯穿智能体技术研发、应用部署与推广的全过程，切实防范系统性风险。坚持规范有序，适应智能体技术演进规律，构建与现有政策法规衔接顺畅、行业自律自治、底线红线清晰的治理体系，有序推进智能体落地应用。坚持创新驱动，加强理论创新、技术创新、工程创新联动，体系化突破智能体关键技术，完善政产学研用协同机制，构建开放共享的智能体生态，提升产业创新活力。坚持应用牵引，重点围绕科学研究、产业发展、提振消费、民生福祉、社会治理等实际需求，发挥典型应用场景示范效应，先易后难、循序渐进，促进智能体技术验证、产品迭代、应用落地。
二、夯实发展基础
夯实技术底座，健全标准体系，降低智能体研发、适配、应用门槛，为丰富智能体产品及服务奠定基础。
（一）完善技术底座
1.强化基础技术研发。持续提升通用基础模型性能，支持行业发展细分领域专用模型，形成适应不同场景和设备的模型产品矩阵。面向智能体训练与运行，提升高质量数据集供给能力。加强智能体任务理解、任务规划、工具使用、长期记忆、互认互通、群体协同等技术攻关，提升智能体泛化能力。
2.完善智能体工具链。开展智能体底层框架研究，加快研发感知、记忆、决策、交互、执行等关键组件，完善智能体研发、测试、部署、运维等工具链。发展对抗样本检测、行为异常检测等安全与治理工具，提升对智能体非合规行为的发现、干预、阻断、恢复能力。
（二）构建标准协议
3.建立智能体标准体系。制定智能体标准化工作指导文件，形成智能体标准框架，系统布局关键技术、重要产品、数据交换、应用场景、质量评测、安全保障、可信认证等标准体系，加快制定智能体与软件工具、应用服务、硬件外设接口等基础标准。加强智能体互联协议（AIP）等智能体互联关键国家标准、行业标准的推广应用。支持医疗、交通、媒体、公共安全等领域制定强制性标准。鼓励企业按照相关标准研发产品服务，提升智能体规范性。积极参与国际标准制定。
4.布局发展智能互联网。研究建立智能互联网体系架构，探索建立智能体注册平台，提供智能体数字身份管理、检索发现、能力声明等服务，支持开发者、部署方式、接口协议、合规认证等信息查询和管理。提升多智能体协同能力，研究智能体身份标识、可信互联、合规支付、安全防护、冲突解决等基础技术。发挥互联网协议第六版（IPv6）技术优势，提升智能体端到端通信能力。探索建立智能互联网监测指标体系。
三、守牢安全底线
坚持以人为本、智能向善、多元共治、安全稳妥，营造规范发展、鼓励创新的制度环境，促进智能体健康有序发展。
（一）明确产品准则
5.完善政策法规和伦理规范。加快研究智能体相关政策法规及伦理规范，发挥专业机构内容资源和审核把关优势，确保智能体行为符合法律法规及主流价值观。防止智能体利用数据优势、人格化技术实施传播不良价值观、算法压榨等行为，防范未成年人、老年人沉迷成瘾、情感依赖等风险。做好与人工智能伦理审查等制度衔接。
6.明确决策权限。在遵守法律法规、尊重社会公德和伦理规范前提下，厘清仅限用户本人决策、需由用户授权决策和智能体自主决策等各种决策方式的合理边界及所需权限。确保用户对智能体自主决策享有知情权和最终决策权，智能体执行操作不得超出用户授权范围。
7.加强行为管控。发展规则内嵌、行为围栏等技术，确保智能体在公共场所、隐私场所、专门场所等的行为合法合规。探索利用区块链等技术，建立重要应用场景智能体行为可验证、可追溯机制，防范智能体不当行为引发重大风险。
（二）防范安全风险
8.提升内生安全能力。研究智能体数据安全、个人信息保护、密码防护、攻击检测、权限管理、行为控制等安全技术，提升智能体系统安全保障能力，防范数据投毒、隐私泄露、算法篡改、系统漏洞、运行失控等安全风险。研究智能体安全检测技术，探索建立智能体安全评估体系。
9.加强供应链安全。制定智能体开发、部署、应用、维护等全周期安全规范，加强模型接入、应用程序接口调用、扩展工具使用等环节安全管理。探索建立智能体供应链安全信息共享和预警机制，及时发布风险提示，提升安全保障能力。
10.化解应用衍生风险。完善智能体常态化风险识别、预警及干预机制，强化人机协同审核、拦截阻断等风险处置能力，防范系统性安全风险。强化智能体应用安全管理，避免智能体被用于自动化攻击、隐私侵犯、虚假信息生成传播、网络诈骗等违法犯罪行为。
（三）完善治理体系
11.构建分类分级治理框架。根据应用场景和潜在影响，审慎稳妥开展智能体分级治理。对于敏感领域及重点行业，由网信部门联合行业主管部门确定开放场景，根据相关法律法规、监管要求和安全防护标准，实行备案、检测、问题产品召回等管理措施。对于部分生活娱乐、日常办公等低风险领域，完善智能体评估测试工具，通过合规自测、信息报告、分发平台管理、行业自律等实现高效治理。
12.健全合规服务体系。强化智能体风险监测预警、检测评估、咨询、认证等专业服务供给，引导行业积极研发智能体监测工具。开展智能体功能、性能、质量、合规等第三方评测服务，推动认证与检测结果互通互认，为用户选择智能体提供参考。编制并发布智能体技术及应用成熟度报告，为产业研发应用提供参考。
（四）强化行业自律
13.引导行业加强自律。鼓励行业组织、主要企业联合制定行业自律规则，明确智能体功能合规、算法治理、知识产权保护、公平竞争等规范细则。指导智能体开发平台、分发平台、服务提供者建立公平合理的平台规则、用户服务协议及隐私政策，明确供需双方权责，保障产业健康发展。加强智能体应用风险宣传教育，提升用户安全意识。
14.探索信用评价机制。指导行业组织建立智能体市场主体自愿参与的信用评价机制，对于技术滥用、诱导消费、虚假宣传、隐瞒缺陷信息等行为进行信用评价，依法依规开展失信惩戒。引导智能体开发者、开发平台、分发平台、服务提供者等参与信用评价，共同营造良好发展环境。
四、强化应用牵引
积极稳妥推动智能体典型场景应用，牵引技术产品优化提升，探索形成可复制、可推广的智能体落地应用模式。
（一）科学研究
15.科研探索。研发理论推演、模拟仿真等智能体，挖掘潜在技术路径。强化智能体信息关联整合、知识体系构建等能力，提升自然科学、哲学社会科学研究发现能力。促进智能体与科学仪器、实验平台融合，实现方案设计、实验操作、数据处理、结果分析等全流程智能化。
16.研发辅助。发展软件开发智能体，提升需求分析、架构设计、代码生成与测试等全流程开发能力。促进智能体与计算机辅助设计（CAD）、计算机辅助工程（CAE）等软件结合，提供设计方案生成、仿真验证、参数调优等功能。
（二）产业发展
17.智能制造。研发生产管理智能体，动态优化生产排程、资源分配和工序衔接，推动智能体在工业互联网领域应用，提升企业精细化管理水平。提升智能体工艺参数优化、加工精度检测、产品缺陷识别等能力，促进智能体与数控机床、工业机器人、自动化产线等融合，促进提质增效降本。
18.能源资源。研发大气、水体、土壤、噪声等环境要素感知智能体，提升自然灾害、环境污染等风险预警能力。依托智能体强化对国土空间资源全周期管理能力。依托智能体实现能源、金属等矿产资源高效勘探。发展电力调度、用电监测、电网维护等智能体，提升电力资源使用效率。
19.交通运输。研发交通安全监管、应急指挥调度等智能体，提升违章违规行为识别、交通基础设施风险预警、重点车辆（船舶）监管、事故快速响应等能力。优化交通监测调度智能体性能，发展交通载运工具管控智能体，提升路网、水网、空域的通行效率。
20.农业生产。研发农业服务智能体，开展农技指导、病虫害诊断与防治等服务。推动智能体在种植养殖、高效育种等环节应用，推进农业智能化转型。推动智能体与智能农机具、智慧大棚、农业服务平台融合，提升农业生产效率。
21.金融服务。研发金融风控智能体，提升信贷审批、交易监控、账户安全等环节风险识别能力。完善智能体异常检测、合规审计功能，提升信贷违约预测、信用卡盗刷拦截、反洗钱监测等能力。
（三）提振消费
22.终端应用。推动智能体赋能互联网应用及服务，优化在线购物、出行导航、生活缴费、日常办公等服务体验。推动智能体与手机、电脑、汽车、家居、可穿戴、消费级机器人等终端设备协同发展，提升跨应用、跨设备任务完成能力。
23.文化旅游。研发文学、音乐、绘画、视听、演艺等内容创作智能体，促进优秀文化传播推广。发展智能导览、多语种翻译、适老适残等旅游服务智能体，提升旅游服务水平。
24.商业服务。提升智能体客服能力，提供7×24小时咨询、预约、售后等服务。发展导引、清洁、仓储、配售等具身智能体，提升餐饮、零售、住宿、物流等商业场所的运营效率。探索通过具身智能体提供低成本家政、养老、托育、助残等服务。
（四）民生福祉
25.教育教学。探索课件生成、作业批改、学情分析等智能体，提高教师工作效率。依托智能体开展个性化学习方案制定，完善智能导学、答疑辅导、虚拟助教等功能。支持在线教育平台研发智能体，提供终身学习服务。
26.医疗健康。提升医学影像分析、疾病诊断推理、定制化诊疗方案生成等医疗辅助智能体性能，探索药品管理、手术排程、病历管理等智能体，提升医疗服务效率。稳步发展预问诊、报告解析等智能体，提升患者体验。
27.人力资源。探索智能体在就业促进、技术技能人才培养评价、劳动关系公共服务等领域应用，提升就业服务能力。发展社会保险、劳动争议仲裁、欠薪治理等智能体，保障劳动者合法权益。
28.信息服务。探索智能体在网络内容建设管理中的应用，鼓励信息发布部门和内容传播平台研发用户分析、选题策划、采编加工、分发推荐、智能审核、舆论引导、情绪疏导、实时翻译等智能体，实现多模态信息、跨领域信息的高效整合。
（五）社会治理
29.政务服务。探索事项辅助审批智能体，推动政务审批流程智能化。发展政策咨询智能体，提供全天在线的政务咨询、流程指引等服务。探索主动推送适配政策、服务提醒及办理指南，加快从“人找服务”向“服务找人”转型。
30.司法服务。探索全流程办案辅助智能体，提升案件材料梳理、案件信息录入、证据审查、辅助法律文书生成等能力。发展法律宣传、法律咨询、法律监督等智能体，为群众提供高效便捷的在线司法服务。
31.公共安全。探索监测预警、应急处置、救援调度、协同治理等智能体，提升安全生产监管和防灾减灾救灾等能力。提升智能体异常行为识别、潜在威胁预警、动态防控处理能力，维护公共安全。推动具身智能体在灾害救援、安防巡检、危险品处置等领域落地应用。
32.城市治理。探索智能体在城市规划、建设与治理环节应用，支撑智能建造、房屋管理、城市基础设施安全运行等工作，提升城市治理专业化水平，提升城市人居环境质量。
33.招标投标。探索招标投标智能体，实现招标投标活动全链路智慧管理，保障全过程规范高效。提升招标投标交易、服务和监管的智慧化水平，实现招标科学合理、评标公平公正和监管穿透高效。
五、建设创新生态
畅通供需渠道，促进研发侧、需求侧高水平互动，形成市场牵引、内驱发展的智能体产业生态。
（一）促进产业合作
34.培育开源创新力量。引导国内人工智能开源社区加强智能体布局，开展智能体与开源芯片、开源操作系统、开源大模型兼容适配。引导企业、高校、科研机构积极参与智能体框架、交互接口、工具链等开源项目，推动技术体系融通发展，加快提升国际影响力。
35.搭建产业协作平台。发挥智能体相关生态联盟、技术验证实验室等产业协作平台的作用，协同产业链上下游开展智能体共性技术研发、标准制定、评估认证等工作，开展智能体技术与产业应用复合人才培养。引导互联网应用、智能终端等领域企业共建生态，探索建立互利共赢的合作模式。
（二）强化应用推广
36.构建应用推广渠道。推动建立智能体软件商店、行业供需信息发布平台，引导智能体企业积极发布产品，形成集聚效应。开展智能体应用供需对接活动，采取公开招标、揭榜挂帅等方式吸引智能体企业定制化开发相应产品。引导整机、软件等企业基于智能体研发产品和服务，培育用户市场。
37.推进重点场景开放。推动重点领域开放智能体应用场景，在产业集聚区、重点行业、重点领域开展智能体应用试点，打造一批具有引领作用的示范项目。发展市场化、专业化的智能体技术转化服务机构，探索智能体应用场景，提升技术成果转化效率。促进行业数据共享开放，支撑重点场景智能体训练部署。
38.积极培育全球生态。依托世界人工智能大会、世界互联网大会等国际平台，交流展示智能体技术创新成果。推动终端设备、软件企业适配智能体，引导相关企业做好海外合规建设，推动智能体适应当地法律法规和文化习俗。
六、保障措施
国家网信办、国家发展改革委、工业和信息化部会同有关方面加强统筹谋划，强化资源整合和力量协同，完善配套政策，形成工作合力，推动重点任务落实落地。建立并完善智能体发展评价指标体系，加强智能体规范应用与创新发展的监测评估、滚动实施和动态调整。
PART II: ENGLISH TRANSLATION
Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents
An intelligent agent is an intelligent system possessing the capacities for autonomous perception, memory, decision-making, interaction, and execution, and constitutes an important form of artificial intelligence products and services. With the rapid development of new-generation artificial intelligence technologies such as large models, intelligent agents are accelerating their deep integration with cyberspace and the physical world, profoundly transforming the modes of human production, daily life, and social governance. In order to implement the State Council’s Opinions on Deeply Implementing the “Artificial Intelligence Plus” Initiative, and to promote the standardized application and innovative development of intelligent agents, these Implementation Opinions are hereby formulated.
I. Fundamental Principles
Guided by the objectives of advancing scientific and technological innovation, enhancing governance capacity, building an industrial ecosystem, and improving the people’s well-being, the following principles shall be upheld: Uphold safety and controllability. The safety, reliability, and trustworthiness of intelligent agents shall be maintained as baseline requirements for development, integrated throughout the entire process of intelligent agent technology research and development, application deployment, and promotion, and systemic risks shall be effectively prevented. Uphold orderly regulation. In accordance with the evolutionary patterns of intelligent agent technology, a governance system shall be constructed that connects smoothly with existing policies and regulations, fosters industry self-discipline and self-governance, and establishes clear baseline and red-line boundaries, so as to advance the implementation of intelligent agents in an orderly manner. Uphold innovation-driven development. Theoretical innovation, technological innovation, and engineering innovation shall be strengthened in a coordinated manner; systematic breakthroughs in key intelligent agent technologies shall be pursued; mechanisms for collaboration among government, industry, academia, research, and end-users shall be improved; and an open and shared intelligent agent ecosystem shall be built to enhance industrial innovation vitality. Uphold application-led advancement. Focused primarily on the practical needs of scientific research, industrial development, consumer stimulation, people’s well-being, and social governance, the demonstrative effects of representative application scenarios shall be leveraged, proceeding from easier to more difficult tasks in a step-by-step manner, to promote the verification of intelligent agent technologies, product iteration, and application implementation.
II. Consolidating the Development Foundation
The technological base shall be consolidated and the standards system shall be strengthened, thereby lowering the barriers to intelligent agent research and development, adaptation, and application, and laying the foundation for enriching intelligent agent products and services.
(1) Improving the Technological Base
1. Strengthening foundational technology research and development. The performance of general-purpose foundational models shall be continuously enhanced; the development of specialized models in segmented industry domains shall be supported, forming a model product matrix adapted to different scenarios and devices. Oriented toward intelligent agent training and operation, the supply capacity of high-quality datasets shall be enhanced. Technical research on intelligent agent task comprehension, task planning, tool usage, long-term memory, mutual recognition and interoperability, and swarm collaboration shall be strengthened to enhance the generalization capabilities of intelligent agents.
2. Improving the intelligent agent toolchain. Research on intelligent agent underlying frameworks shall be conducted; the development of key components for perception, memory, decision-making, interaction, and execution shall be accelerated; and the toolchain for intelligent agent research and development, testing, deployment, and operations and maintenance shall be improved. Security and governance tools such as adversarial sample detection and behavioral anomaly detection shall be developed to enhance the capacity for discovering, intervening in, blocking, and recovering from non-compliant intelligent agent behaviors.
(2) Building Standards and Protocols
3. Establishing an intelligent agent standards system. Guidance documents for intelligent agent standardization work shall be formulated, forming an intelligent agent standards framework that systematically arranges standards systems for key technologies, important products, data exchange, application scenarios, quality evaluation, security assurance, and trustworthy certification, while accelerating the formulation of foundational standards for intelligent agent interfaces with software tools, application services, and hardware peripherals. The promotion and application of key national and industry standards such as the Agent Interconnection Protocol (AIP) shall be strengthened. The formulation of mandatory standards in fields such as healthcare, transportation, media, and public safety shall be supported. Enterprises shall be encouraged to develop products and services in accordance with relevant standards to enhance intelligent agent standardization. Active participation in international standards formulation shall be pursued.
4. Planning and developing the intelligent interconnection network. Research shall be conducted to establish the architecture of the intelligent interconnection network; the establishment of intelligent agent registration platforms shall be explored, providing services such as intelligent agent digital identity management, search and discovery, and capability declaration, supporting the querying and management of information on developers, deployment methods, interface protocols, and compliance certifications. Multi-agent collaboration capabilities shall be enhanced; research shall be conducted on foundational technologies for intelligent agent identity identification, trusted interconnection, compliant payment, security protection, and conflict resolution. The technological advantages of Internet Protocol version 6 (IPv6) shall be leveraged to enhance end-to-end communication capabilities of intelligent agents. The establishment of an intelligent interconnection network monitoring indicator system shall be explored.
III. Safeguarding the Security Baseline
Adhering to the principles of human-centricity, intelligence for good, pluralistic co-governance, and safety and prudence, a regulatory environment that promotes standardized development and encourages innovation shall be cultivated to promote the healthy and orderly development of intelligent agents.
(1) Clarifying Product Standards
5. Improving policies, regulations, and ethical norms. Research on intelligent agent-related policies, regulations, and ethical norms shall be accelerated; the content resources and review and gatekeeping advantages of professional institutions shall be leveraged to ensure that intelligent agent behavior conforms to laws, regulations, and mainstream values. Intelligent agents shall be prevented from utilizing data advantages and personification technologies to disseminate unwholesome values, engage in algorithmic exploitation, and other such behaviors; and risks such as addiction, emotional dependency among minors and the elderly shall be guarded against. Proper alignment with systems such as AI ethics review shall be maintained.
6. Clarifying decision-making authority. On the premise of compliance with laws and regulations, and respect for social morality and ethical norms, the reasonable boundaries and required permissions for various decision-making modalities—including decisions restricted to the user alone, decisions requiring user authorization, and autonomous decisions by intelligent agents—shall be delineated. Users’ rights to be informed of and to exercise final authority over intelligent agent autonomous decisions shall be ensured, and the execution of operations by intelligent agents shall not exceed the scope of user authorization.
7. Strengthening behavioral controls. Technologies such as rule embedding and behavioral guardrails shall be developed to ensure that the behavior of intelligent agents in public spaces, private spaces, and specialized spaces is lawful and compliant. The use of technologies such as blockchain shall be explored to establish verifiable and traceable mechanisms for intelligent agent behavior in important application scenarios, guarding against major risks arising from improper intelligent agent behavior.
(2) Preventing Security Risks
8. Enhancing intrinsic security capabilities. Research shall be conducted on security technologies for intelligent agent data security, personal information protection, cryptographic defense, attack detection, permission management, and behavioral control, enhancing the security assurance capabilities of intelligent agent systems, and guarding against security risks such as data poisoning, privacy leakage, algorithm tampering, system vulnerabilities, and loss of operational control. Research on intelligent agent security detection technologies shall be conducted, and the establishment of an intelligent agent security assessment system shall be explored.
9. Strengthening supply chain security. Full-lifecycle security specifications for intelligent agent development, deployment, application, and maintenance shall be formulated, and security management of links such as model access, application programming interface calls, and extended tool usage shall be strengthened. The establishment of intelligent agent supply chain security information sharing and early warning mechanisms shall be explored, timely risk advisories shall be issued, and security assurance capabilities shall be enhanced.
10. Resolving application-derived risks. Routine risk identification, early warning, and intervention mechanisms for intelligent agents shall be improved; risk handling capabilities such as human-machine collaborative review and interception and blocking shall be strengthened to guard against systemic security risks. The security management of intelligent agent applications shall be reinforced to prevent intelligent agents from being used for automated attacks, privacy violations, generation and dissemination of false information, cyber fraud, and other illegal and criminal activities.
(3) Improving the Governance System
11. Building a classified and tiered governance framework. In accordance with application scenarios and potential impacts, classified and tiered governance of intelligent agents shall be carried out in a prudent and steady manner. For sensitive domains and key industries, cyberspace administration departments shall jointly determine open scenarios with sectoral regulators, implementing management measures such as registration, testing, and defective product recall in accordance with relevant laws and regulations, regulatory requirements, and security protection standards. For certain low-risk domains such as lifestyle and entertainment and daily office work, intelligent agent assessment and testing tools shall be improved, achieving efficient governance through compliance self-testing, information reporting, distribution platform management, and industry self-discipline.
12. Strengthening the compliance services system. The supply of professional services such as intelligent agent risk monitoring and early warning, testing and assessment, consultation, and certification shall be strengthened, and the industry shall be guided to actively develop intelligent agent monitoring tools. Third-party evaluation services for intelligent agent functionality, performance, quality, and compliance shall be conducted; the mutual recognition and interoperability of certification and testing results shall be promoted to provide users with reference for selecting intelligent agents. Intelligent agent technology and application maturity reports shall be compiled and published to provide reference for industrial research, development, and application.
(4) Strengthening Industry Self-Discipline
13. Guiding the industry to strengthen self-discipline. Industry organizations and major enterprises shall be encouraged to jointly formulate industry self-discipline rules, clarifying detailed rules for intelligent agent functional compliance, algorithm governance, intellectual property protection, and fair competition. Intelligent agent development platforms, distribution platforms, and service providers shall be guided to establish fair and reasonable platform rules, user service agreements, and privacy policies, clarifying the rights and responsibilities of both supply and demand sides to safeguard healthy industrial development. Public education on intelligent agent application risks shall be strengthened to enhance user security awareness.
14. Exploring credit evaluation mechanisms. Industry organizations shall be guided to establish voluntary credit evaluation mechanisms for intelligent agent market participants, conducting credit evaluations for behaviors such as technology misuse, consumer inducement, false advertising, and concealment of defect information, and carrying out sanctions for credibility loss in accordance with law and regulation. Intelligent agent developers, development platforms, distribution platforms, service providers, and others shall be guided to participate in credit evaluation to jointly cultivate a favorable development environment.
IV. Strengthening Application-Led Advancement
The application of intelligent agents in representative scenarios shall be promoted in an active and prudent manner, guiding the optimization and improvement of technological products, and exploring replicable and scalable models for intelligent agent implementation.
(1) Scientific Research
15. Scientific exploration. Intelligent agents for theoretical derivation, simulation, and other functions shall be developed to uncover potential technological pathways. The capabilities of intelligent agents in information correlation and integration, and knowledge system construction shall be strengthened to enhance the discovery capacity of natural sciences and philosophy and social sciences research. The integration of intelligent agents with scientific instruments and experimental platforms shall be promoted to achieve full-process intelligent automation including scheme design, experimental operations, data processing, and results analysis.
16. R&D assistance. Software development intelligent agents shall be developed to enhance full-process development capabilities including requirements analysis, architectural design, code generation, and testing. The integration of intelligent agents with computer-aided design (CAD), computer-aided engineering (CAE), and other software shall be promoted, providing functions such as design scheme generation, simulation verification, and parameter optimization.
(2) Industrial Development
17. Intelligent manufacturing. Production management intelligent agents shall be developed to dynamically optimize production scheduling, resource allocation, and process sequencing, promoting the application of intelligent agents in the industrial internet domain and enhancing enterprises’ refined management capabilities. The capabilities of intelligent agents in process parameter optimization, machining precision detection, and product defect identification shall be enhanced, promoting the integration of intelligent agents with CNC machine tools, industrial robots, and automated production lines to improve quality, increase efficiency, and reduce costs.
18. Energy and resources. Intelligent agents for sensing environmental elements such as atmosphere, water bodies, soil, and noise shall be developed to enhance risk early warning capabilities for natural disasters and environmental pollution. The management capacity over the full lifecycle of territorial spatial resources shall be strengthened through intelligent agents. Efficient exploration of energy, metals, and other mineral resources shall be achieved through intelligent agents. Intelligent agents for power dispatch, electricity consumption monitoring, and power grid maintenance shall be developed to enhance the efficiency of power resource utilization.
19. Transportation. Intelligent agents for transportation safety supervision, emergency command and dispatch, and other functions shall be developed, enhancing capabilities in traffic violation identification, transportation infrastructure risk early warning, key vehicle (vessel) supervision, and rapid accident response. The performance of traffic monitoring and dispatch intelligent agents shall be optimized; intelligent agents for traffic carrier control shall be developed to improve the throughput efficiency of road networks, waterway networks, and airspace.
20. Agricultural production. Agricultural service intelligent agents shall be developed to provide services such as agricultural technical guidance and pest and disease diagnosis and prevention. The application of intelligent agents in planting, breeding, efficient seed cultivation, and other processes shall be promoted to advance the intelligent transformation of agriculture. The integration of intelligent agents with smart agricultural machinery, smart greenhouses, and agricultural service platforms shall be promoted to enhance agricultural production efficiency.
21. Financial services. Financial risk management intelligent agents shall be developed to enhance risk identification capabilities in credit approval, transaction monitoring, account security, and other areas. The anomaly detection and compliance auditing functions of intelligent agents shall be improved, enhancing capabilities in credit default prediction, credit card fraud interception, and anti-money laundering monitoring.
(3) Stimulating Consumption
22. Terminal applications. The empowerment of internet applications and services by intelligent agents shall be promoted, optimizing service experiences in online shopping, travel navigation, utility payments, daily office work, and other areas. The coordinated development of intelligent agents with terminal devices such as mobile phones, computers, automobiles, home appliances, wearables, and consumer-grade robots shall be promoted, enhancing cross-application and cross-device task completion capabilities.
23. Culture and tourism. Intelligent agents for content creation in literature, music, painting, audiovisual media, and performing arts shall be developed to promote the dissemination of outstanding culture. Intelligent tour guide, multilingual translation, and elderly- and disability-accessible tourism service intelligent agents shall be developed to enhance the level of tourism services.
24. Commercial services. The customer service capabilities of intelligent agents shall be enhanced to provide 24/7 consultation, appointment, and after-sales services. Embodied intelligent agents for guidance, cleaning, warehousing, and distribution shall be developed to improve the operational efficiency of commercial venues such as restaurants, retail stores, lodging facilities, and logistics centers. The provision of low-cost domestic help, elderly care, childcare, and disability assistance services through embodied intelligent agents shall be explored.
(4) People’s Well-Being
25. Education and teaching. Intelligent agents for courseware generation, homework grading, and learning analytics shall be explored to improve teacher work efficiency. Personalized learning plan development shall be conducted through intelligent agents, improving functions such as intelligent learning guidance, question-and-answer tutoring, and virtual teaching assistants. Online education platforms shall be supported in developing intelligent agents to provide lifelong learning services.
26. Healthcare. The performance of medical assistance intelligent agents for medical image analysis, disease diagnostic reasoning, and customized treatment plan generation shall be enhanced; intelligent agents for pharmaceutical management, surgical scheduling, and medical record management shall be explored to improve healthcare service efficiency. Pre-consultation and report interpretation intelligent agents shall be steadily developed to enhance the patient experience.
27. Human resources. The application of intelligent agents in employment promotion, technical and vocational talent cultivation and evaluation, and labor relations public services shall be explored to enhance employment service capabilities. Intelligent agents for social insurance, labor dispute arbitration, and wage arrears governance shall be developed to safeguard the lawful rights and interests of workers.
28. Information services. The application of intelligent agents in online content construction and management shall be explored; information publishing departments and content distribution platforms shall be encouraged to develop intelligent agents for user analysis, topic planning, content editing and processing, distribution and recommendation, intelligent review, public opinion guidance, emotional counseling, and real-time translation, achieving the efficient integration of multimodal and cross-domain information.
(5) Social Governance
29. Government services. Intelligent agents for assisting in the review and approval of government matters shall be explored, promoting the intelligent transformation of government approval processes. Policy consultation intelligent agents shall be developed to provide round-the-clock online government consultation, process guidance, and other services. The proactive delivery of tailored policies, service reminders, and processing guides shall be explored, accelerating the transformation from “people seeking services” to “services finding people.”
30. Judicial services. Full-process case-handling assistance intelligent agents shall be explored, enhancing capabilities in case material organization, case information entry, evidence review, and assisted legal document generation. Intelligent agents for legal publicity, legal consultation, and legal supervision shall be developed to provide the public with efficient and convenient online judicial services.
31. Public safety. Intelligent agents for monitoring and early warning, emergency response, rescue dispatch, and collaborative governance shall be explored to enhance capabilities in work safety supervision and disaster prevention, mitigation, and relief. The capabilities of intelligent agents in anomalous behavior identification, potential threat early warning, and dynamic prevention and control shall be enhanced to maintain public safety. The implementation of embodied intelligent agents in disaster rescue, security patrol and inspection, and hazardous materials handling shall be promoted.
32. Urban governance. The application of intelligent agents in urban planning, construction, and governance shall be explored, supporting intelligent construction, housing management, and the safe operation of urban infrastructure, enhancing the professionalization of urban governance and improving the quality of urban living environments.
33. Bidding and tendering. Intelligent agents for bidding and tendering shall be explored, achieving full-chain intelligent management of bidding and tendering activities and ensuring standardized and efficient processes throughout. The level of intelligence in bidding and tendering transactions, services, and supervision shall be enhanced, achieving scientifically sound bidding, fair and impartial bid evaluation, and penetrating and efficient supervision.
V. Building an Innovation Ecosystem
Supply and demand channels shall be facilitated, promoting high-level interaction between the R&D side and the demand side, forming a market-driven, internally motivated intelligent agent industrial ecosystem.
(1) Promoting Industrial Cooperation
34. Cultivating open-source innovation forces. Domestic AI open-source communities shall be guided to strengthen their intelligent agent deployment; compatibility and adaptation of intelligent agents with open-source chips, open-source operating systems, and open-source large models shall be pursued. Enterprises, universities, and research institutions shall be guided to actively participate in open-source projects for intelligent agent frameworks, interaction interfaces, and toolchains, promoting the integrated development of technology systems and accelerating the enhancement of international influence.
35. Building industrial collaboration platforms. The role of industrial collaboration platforms such as intelligent agent-related ecosystem alliances and technology verification laboratories shall be leveraged, coordinating upstream and downstream participants in the industrial chain to conduct common technology R&D, standards formulation, and assessment and certification for intelligent agents, and cultivating talent with combined expertise in intelligent agent technology and industrial applications. Enterprises in internet applications, intelligent terminals, and other domains shall be guided to co-build ecosystems, exploring the establishment of mutually beneficial cooperation models.
(2) Strengthening Application Promotion
36. Building application promotion channels. The establishment of intelligent agent app stores and industry supply-and-demand information publishing platforms shall be promoted, guiding intelligent agent enterprises to actively publish products and creating aggregation effects. Intelligent agent application supply-and-demand matching activities shall be conducted, using methods such as open bidding and competitive challenges to attract intelligent agent enterprises to develop customized products. Enterprises in complete systems, software, and other fields shall be guided to develop products and services based on intelligent agents, cultivating the user market.
37. Advancing the opening of key scenarios. The opening of intelligent agent application scenarios in key domains shall be promoted; intelligent agent application pilots shall be conducted in industrial clusters, key industries, and key domains, creating a number of demonstration projects with leading effects. Market-oriented and professional intelligent agent technology transformation service institutions shall be developed; intelligent agent application scenarios shall be explored to improve the efficiency of technology commercialization. Industry data sharing and openness shall be promoted to support the training and deployment of intelligent agents for key scenarios.
38. Actively cultivating a global ecosystem. Leveraging international platforms such as the World Artificial Intelligence Conference and the World Internet Conference, intelligent agent technology innovation achievements shall be exchanged and showcased. Terminal equipment and software enterprises shall be encouraged to adapt to intelligent agents; relevant enterprises shall be guided to establish overseas compliance frameworks, promoting the adaptation of intelligent agents to local laws, regulations, and cultural customs.
VI. Safeguard Measures
The Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology, together with relevant parties, shall strengthen overall planning and coordination, reinforce resource integration and collaborative efforts, improve supporting policies, form a concerted working force, and promote the implementation of key tasks. An evaluation indicator system for intelligent agent development shall be established and improved, and the monitoring, assessment, rolling implementation, and dynamic adjustment of the standardized application and innovative development of intelligent agents shall be strengthened.
PART III: ANALYSIS — THE IMPLEMENTATION OPINIONS READ THROUGH THE WORK OF LARRY CATÁ BACKER
The Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents provide a concrete site at which I can test the thesis developed in my working paper, ‘“Modernization” (现代化) and Chinese Constitutionalism’s “Lifeworld”: Constitutionalism as the Institutionalized Phenomenology of Modernization — A Semiotic-Phenomenological Analysis of Modernization Discourse in Chinese Marxist-Leninist Political-Legal Theory’ (Working Paper Draft, August 2026) (the “Modernization paper”). I read the Implementation Opinions not as an isolated AI policy document but as a manifestation of Chinese Marxist-Leninist modernization theory under New Era structures, aligned with the 3rd and 4th Plenums’ focus on advancing modernization, high-quality development, and new quality productive forces. What follows is therefore an analysis in my own first-person voice: I use the Implementation Opinions to extend, specify, and test my account of modernization as the constitutional lifeworld of Chinese governance.
The Modernizing Lebenswelt and Intelligent Agents
In the Modernization paper, I argue that modernization functions as the Lebenswelt—the pre-given, intersubjective horizon of meaning, in Husserl’s sense—within which Chinese constitutionalism constitutes itself. Modernization is not simply a policy objective placed alongside socialism, Party leadership, reform, governance, security, or national rejuvenation; it is the practical and semantic horizon that makes their relations intelligible.
This is the metasignifier thesis. Modernization organizes other political signs—socialism, Party leadership, reform, opening, law, governance, common prosperity, ecological civilization, security, and rejuvenation—without emptying them of content. It mediates among them, assigning each a historically specific place within one project while preserving its distinct genealogy and institutional work. The Implementation Opinions’ cross-domain architecture can therefore be read as a semiotic operation, not merely as a list of AI applications.
I read the Implementation Opinions as the newest deposit within the sedimentary ontology I describe in the Modernization paper—whose nine layers are: backwardness; revolution; socialist construction; industrialization and the Four Modernizations; reform and opening; the socialist market economy; governance modernization; Chinese modernization; and national rejuvenation. The Implementation Opinions do not constitute a distinct epoch-scale layer comparable to these; they represent a qualitative thickening of the governance-modernization layer, because the state’s own instruments of observation, coordination, administration, and control become objects of modernization themselves.
The 3rd and 4th Plenums: High-Quality Development and New Quality Productive Forces
The 3rd Plenum Decision, “Further Deepening Reform Comprehensively to Advance Chinese Modernization” (July 2024), places reform explicitly in a subordinate but constitutive relation to modernization: reform is the means by which the modernizing project advances. The 4th Plenum Recommendations for the 15th Five-Year Plan (October 2025) extend that architecture across industrial modernization, scientific and technological self-reliance, new quality productive forces, the domestic market, rural and defense modernization, green transition, national security, and Party governance. The breadth of these instruments shows the metasignifier at work across the planning cycle.
The theoretical license for this emphasis on quality and comprehensive transformation lies in the 19th Congress’s reformulation of the principal contradiction—from the contradiction between the people’s growing material and cultural needs and backward social production to the contradiction between unbalanced and inadequate development and the people’s ever-growing needs for a better life. The shift acknowledges that the central problem is no longer aggregate productive insufficiency alone. It reindexes modernization toward balance, adequacy, quality, responsiveness, and the capacity of institutions to address differentiated needs.
“High-quality development” (高质量发展) bridges abstract modernization theory and concrete governance by translating those qualitative requirements into innovation, sustainability, equity, and institutional performance. “New quality productive forces” (新质生产力) is the latest specification of productive-forces development within that bridge: innovation-driven, knowledge-intensive, technologically advanced capacity that transforms production across sectors. Intelligent agents fall within the domain this formulation addresses. They reorganize scientific research, industrial production, services, public administration, and social governance through systems capable of autonomous perception, decision-making, tool use, and execution—though the concept of new quality productive forces is broader than intelligent agents alone and encompasses other frontier technologies as well.
The Five-Year Plan supplies the system’s temporal rhythm: it periodizes the modernizing project, converts abstract commitments into time-bound targets, and provides the evaluative framework against which institutional adequacy is measured. The Implementation Opinions should therefore be read within the temporal frame of the 15th Five-Year Plan, not as a free-standing technology policy.
AI as Strategic Infrastructure
In my 2026 comparative lecture series at East China University of Political Science and Law, I describe China as treating AI “as strategic infrastructure for socialist modernization, platform governance, data security, public opinion management, and national rejuvenation,” in contrast to the EU’s rights-and-risk framing or the U.S.’s fragmented sectoral approach.[1] The Implementation Opinions are broadly consistent with this characterization. They do not confine intelligent agents to a product-safety framework; they envision deployment across scientific research, industrial production, energy, transportation, agriculture, financial services, education, healthcare, government approvals, judicial case-handling, public safety, and urban governance (Sections 15–33). That cross-sectoral scope is the metasignifier at work: heterogeneous domains are organized as differentiated sites within one modernizing project. This range supports my earlier claim that Chinese AI governance is “not reducible to privacy or product safety” but is “broader and more integrated across cybersecurity, algorithmic recommendation, deep synthesis, generative AI, industrial policy, and national security.”[2]
I do not mean that cross-sectoral deployment is uniquely Chinese; many national AI strategies envision it. What distinguishes the Chinese approach in my prior framing is the integration of AI governance with the Party-state’s political project of modernization and the CPC’s leadership role. Read through historical materialism, intelligent agents represent one of the leading technologies through which the vanguard party claims to advance productive-forces development by means of conscious, Party-directed transformation of production, administration, and social coordination. The Implementation Opinions’ preamble, which grounds the document in the State Council’s “AI Plus” initiative, and Section VI, which names the Cyberspace Administration of China, the NDRC, and MIIT as coordinating bodies, reflect this Party-state coordination structure.[3][4]
The Development-Security Dialectic
A concept central to my Shanghai Lecture 5—“The Guiding State: The Chinese Approach”—and absent from the earlier version of this analysis is the “development-security dialectic.” I read it through Marxist-Leninist theory, in which the vanguard party claims the capacity to direct technological transformation as a conscious, planned process rather than leave it to an autonomous developmental logic. In my formulation, modernization operates as both “engine and vessel for the rationalization of the state.”[5] Chinese policy “consistently combines two imperatives: AI must advance economic growth, industrial upgrading, and technological capacity (development), while also remaining controllable, secure, and aligned with national security, social stability, and ideological requirements (security).”[6] These imperatives are not treated as competing goals in a liberal balancing exercise. Within the Modernization paper’s framework, development is pursued through mechanisms that embed security, and security is designed to enable sustainable development under state supervision.
I see the Implementation Opinions displaying this dialectic throughout. Its four fundamental principles pair innovation-driven development and application-led advancement with safety-and-controllability and orderly regulation. The pairing also follows the Leninist developmental grammar I trace from Lenin’s electrification thesis through Chinese modernization: organized political direction mobilizes technical capacity, technical capacity expands productive forces, and security protects the conditions for continued transformation. Section II (“Consolidating the Development Foundation”) and Section III (“Safeguarding the Security Baseline”) are structurally parallel, reflecting the paired character of the dialectic. Whether the Implementation Opinions sustain this pairing—or whether security imperatives predominate in practice—is a question the document cannot answer for itself, and one I leave open for the moment.[7]
The Coordinated Regulatory Architecture and Specific Chinese Provisions
My Lecture 5 identifies the specific legal and policy instruments that constitute China’s AI governance architecture. These are described “not as components of a unified regulatory code, but as elements of a coordinated architecture through which party-state priorities are operationalized across the AI stack.”[8] These include:
— The New Generation Artificial Intelligence Development Plan (2017)
— The Cybersecurity Law
— The Data Security Law
— The Personal Information Protection Law
— The Provisions on Algorithmic Recommendation
— The Provisions on Deep Synthesis Internet Information Services
— The Interim Measures for Generative Artificial Intelligence Services
One might approach this as “coordination under Party leadership rather than integration into a single legal code.”[9] The Implementation Opinions add a new layer to this architecture — one focused specifically on intelligent agents as a distinct regulatory category. Unlike the existing instruments, which regulate data, algorithms, synthetic content, and generative AI outputs respectively, the Implementation Opinions address autonomous, interactive, tool-using systems that may engage with all of those regulatory layers simultaneously. This raises the question — which remains unexamined, since the Implementation Opinions are not among the instruments I have analyzed — of how agent-specific regulation will interact with the existing layered instruments, particularly where an intelligent agent simultaneously triggers obligations under algorithmic recommendation rules, generative AI measures, and data security law. The Implementation Opinions do not resolve this interaction; they assume coordination without specifying its mechanics.
Public Opinion Attributes and Social Mobilization Capacity
A concept I highlight in Lecture 5 that is worth examining in connection with the Implementation Opinions is “public opinion attributes” (舆论属性) and “social mobilization capacity” (社会动员能力). I noted that the Provisions on Algorithmic Recommendation and the Deep Synthesis rules “treat recommender systems and synthetic media as infrastructures of perception and coordination, capable of shaping collective cognition, discourse, and action. Their governance is therefore inseparable from the Party-state’s responsibility for public opinion guidance and social order.”[10]
The Implementation Opinions’ Section 28, on information services, explicitly envisions intelligent agents performing “public opinion guidance” (舆论引导) and “emotional counseling” (情绪疏导) functions. This is a notable provision. It suggests that agents are being conceived not only as tools for productivity and service delivery but as instruments within the Party-state’s apparatus for managing public discourse and sentiment. Whether this provision will trigger the heightened oversight that I would associate with systems having “public opinion attributes” under existing algorithmic recommendation rules is unclear from the document itself. But the conceptual link is worth noting: the Implementation Opinions appear to contemplate agents operating in precisely the space—shaping collective cognition and discourse—that I identify as most distinctively governed under the Chinese approach.
Platforms as Governance Intermediaries
My Lecture 5 emphasizes that in China, “large technology companies function as governance intermediaries. They are responsible for implementing regulatory requirements through system design, content moderation, data management, and compliance processes. This is not simply private self-regulation. It is a form of delegated governance in which platforms operationalize Party-state directives.”[11] In the Modernization paper, I place this insight alongside Foucault’s account of governmentality: platforms are governmental technologies through which a political rationality is translated into conduct.
The Implementation Opinions assign precisely this intermediary role to “development platforms” (开发平台) and “distribution platforms” (分发平台) for intelligent agents. Section 13 directs these platforms to establish “fair and reasonable platform rules, user service agreements and privacy policies” and to define the “rights and responsibilities of both supply and demand sides.” Section 14’s credit evaluation mechanism also runs through these platforms. This is consistent with my broader observation that governance is “distributed across actors but coordinated through political authority.”[12] More importantly, these platform practices are the institutional channels through which macro-level modernization discourse becomes micro-level institutional practice. Registration, platform rules, privacy policies, distribution controls, credit evaluation, and compliance processes make the modernizing imperative circulate capillarily through everyday governance. I therefore ask whether agent distribution platforms will assume the same governance-intermediary burden that search engines and social media platforms already bear under the algorithmic recommendation and content-governance regimes—and what the consequences would be if they resist or prove unable to fulfill that role.
The “Operating System” Thesis
My 2018 paper “And an Algorithm to Bind Them All?” argues that data-driven governance systems represent “an operating system for global normative orders” that entangles with but is not reducible to conventional law.[13] In this 2026 analysis of Xi Jinping’s WAIC keynote, I sought to extend this reading, arguing that China is “not merely proposing rules but building an institutional ‘operating system’ that other governance orders (public and private, domestic and international) can be plugged into, with China occupying the position of systems architect.”[14]
The Implementation Opinions’ Section 4—envisioning an “intelligent interconnection network” with agent registration platforms, digital identity management, capability declarations, and interoperability protocols—resonates with this line of analysis. However, the fit should be stated honestly. My operating-system thesis addresses the displacement of conventional law by data-driven behavioral governance at a polycentric, often global level. The Implementation Opinions’ intelligent interconnection network is, on its face, a more conventional infrastructure-standardization project. Whether it becomes the kind of operating system I theorize depends on how it is implemented and whether the registration, identity, and compliance-certification apparatus evolves into a mechanism that governs agent behavior through its own logic rather than merely serving as a registry. The document leaves that question open.
The MPS Cyberspace Security Measures and the Supervision Architecture
A provision analyzed in August 2026 that bears on the Implementation Opinions is the Ministry of Public Security’s Measures on Cyberspace Security Supervision and Inspection (MPS Order No. 176, effective October 2026). There, these measures are considered as completing “the transition of public security organs’ role from primarily ‘internet police’ under the 2018 framework to a comprehensive cyberspace-security supervisor operating within the modern tripartite (network–data–information) regulatory system.”[15]
The Implementation Opinions’ Section 8 (intrinsic security capabilities), Section 9 (supply chain security), and Section 10 (application-derived risks) do not specify which organs will conduct supervision and inspection, but the existing institutional architecture — including MPS Order No. 176’s inspection powers over network operators, data processors, and personal information handlers — provides the enforcement apparatus within which the Implementation Opinions’ security provisions would presumably operate. The graded/classified inspection system that I describe in the MPS measures mirrors the Implementation Opinions’ own “classified and tiered governance framework” (Section 11). Whether the two instruments will be coordinated or will create overlapping supervisory burdens is an open question.
The NFRA Financial Sector Guidance and Sector-Specific Layering
My analysis of the NFRA’s Guiding Opinions on AI in Banking and Insurance (June 2026) provides a template for understanding how the Implementation Opinions’ general provisions might be operationalized at the sectoral level. There I describe the NFRA guidance as “another layer of a construct that is increasingly densely coordinating set of layers of rules that are meant to operate to guide administrative discretion even as it provides officials their goals.”[16] I note that each such instrument “forms one element” of a system “aligned with systems of administrative inspection and supervision that are meant to provide local and national accountability for decisions actually made, work produced, and results.”
The Implementation Opinions’ Section 21, on financial services, closely parallels the NFRA guidance — both address financial risk management agents, anomaly detection, and compliance auditing. The question is one of layering: the NFRA guidance operates with specificity at the sectoral level (32 measures addressing governance architecture, development, data governance, computing infrastructure, risk management, capability, and oversight); the Implementation Opinions operate at a higher level of generality. How these layers interact — whether the Implementation Opinions add obligations beyond what the NFRA guidance already requires, or merely duplicate them — is not specified.
The Xue Lan Analysis and “Governance Theology”
In August 2026, in an analysis of a keynote by Xue Lan (薛澜), chair of China’s National Expert Committee on AI Governance, I suggested that Xue’s speech as “governance theology — an authoritative account, delivered by the state’s own designated AI-governance theorist, of why the current expansion of PRC administrative control over AI development (the revised Cybersecurity Law, the deep-synthesis and generative-AI rules, the sandbox and certification apparatus) is not merely permissible but sociologically necessary.”[17]
My critique is that what Xue presents “as sociological description is, at the level of the text’s actual referents, a recitation of the current administrative-regulatory apparatus” — that “the theoretical apparatus of co-evolution and large technical systems is doing legitimating work for a specific, already-existing, already-operating body of law — it is not a neutral analytic frame from which policy might be derived; it is a post-hoc justification for policy already made.”
The Implementation Opinions might be read with a similar awareness. The document’s invocation of “application-led advancement” (应用牵引) and its emphasis on building a sociotechnical ecosystem around intelligent agents echoes Xue’s arguments about the need to construct a “sociotechnical system” to realize AI’s potential. Whether the Implementation Opinions’ framework is a genuine governance innovation or a rationalization of Party-state control over a new category of technology is a question my analysis of Xue suggests I should at least ask.
The Instrumentalization Contradiction
In my analysis of the NFRA guidance, I identified what I call “the fundamental issue of instrumentalization and capacity”: “the more autonomous the tech, the greater the risk that the relationship between instrument and its wielders will be reversed, at least in part.”[18] I described two contradictions: “between the leadership of the Party and its capacity to lead,” and “between the techno-instruments through which Party capacity is undertaken and the ability of the Party apparatus to steer, guide, assess, control and utilize these instruments.” Read through the Modernization paper’s recursive structure, these are not merely external risks to governance; they arise because the state’s instruments become objects of modernization.
I see this tension in the Implementation Opinions, but I do not treat it as the document’s sole organizing logic. The document envisions intelligent agents assisting in government approvals (Section 29), judicial case handling (Section 30), public safety (Section 31), and other governance functions. At the same time, it insists on safety and controllability as baseline requirements and mandates human final authority over agent decisions. The risk is recursive: autonomous technology may become sufficiently complex, distributed, or self-directing to exceed the political apparatus’s capacity to steer, assess, control, and utilize it. Within the sedimentary ontology, this tension is a structural feature of the governance-modernization layer becoming reflexive—the state must modernize the very instruments through which it claims to modernize society. The repeated use of “explore” (探索) for governance-adjacent applications, rather than “promote” (推动) or “develop” (发展), may reflect a cautious, experimental approach, although it may also be ordinary bureaucratic hedging. The document does not resolve this ambiguity.
Credit Evaluation and Social Credit
Section 14 of the Implementation Opinions proposes “credit evaluation mechanisms” for intelligent agent market participants—covering technology misuse, consumer inducement, false advertising, and concealment of defect information—with sanctions for credibility loss. I have extensively tracked how social credit is “a functionally differentiated aspect of a much broader undertaking that is meant to develop a Marxist-Leninist structure for the incorporation of tech based governance within the core political project of the CPC.”[19]
The structural resemblance between Section 14’s credit evaluation and broader social credit mechanisms is real: both use reputational assessment and credibility sanctions as governance tools. Yet to be specifically analyzed are intelligent-agent credit evaluation as an extension of social credit. The connection is an inference from my broader framework. Moreover, the Implementation Opinions describe the mechanism as “voluntary” (自愿参与), which complicates any equation with the compulsory dimensions of social credit.
Classified and Tiered Governance
Section 11’s “classified and tiered governance framework” invites comparison with both the EU AI Act’s risk-tiered classification and with my development of a way of examining Chinese regulatory layering. I would resist a simple EU comparison. In my lecture series, I distinguish the EU’s “risk-based, ex ante, lifecycle approach grounded in rights and procedural supervision” from China’s model, which uses “party-state coordination, administrative speed, and integration of AI policy with industrial and security goals.”[20] The Implementation Opinions’ tiered framework is organized not around abstract risk levels but around administrative authority—with the cyberspace administration and sectoral regulators jointly determining which scenarios are open in sensitive domains, and lighter-touch self-governance applying in low-risk domains. This is administrative management rather than risk classification in the EU sense.
Theanalysis of the MPS cyberspace security measures reinforces this point: the Chinese approach uses “graded and classified” (分级分类) governance not primarily as a risk-assessment methodology but as an allocation of administrative responsibility and inspection intensity across different levels of the state apparatus.[21]
The Semiotic Dimension
My approach to Chinese regulatory texts is explicitly semiotic. In my WAIC analysis, I describe attending to how “the speech’s own rhetorical form is already rehearsing, at the level of text, the closure that the institution and the architecture are enacting at the level of governance.”[22] Applied to the Implementation Opinions, I note the document’s four principles and their ordering: safety and controllability first, then orderly regulation, innovation-driven development, and application-led advancement. I therefore ask what work this ordering does—whether it signals a priority structure or merely reflects drafting convention.
Similarly, the concept of “application-led advancement” (应用牵引)—literally “application traction” or “pull”—encodes a specific theory of innovation in which deployment needs drive technological development, consistent with the broader New Era emphasis on new quality productive forces. My analysis of Xue Lan suggests reading such formulations not at face value but as part of the broader “signification matrix” in which New Era doctrinal content operates beneath a borrowed vocabulary of sociotechnical systems theory.[23] Whether this specific reading applies to the Implementation Opinions—a more conventional policy document than Xue’s academic keynote—remains debatable.
The Juridification Cascade
In the Modernization paper, I use “juridification” to describe the cascade through which political theory becomes constitutional purpose, constitutional purpose becomes legislative method, legislative method becomes administrative organization, and administrative organization becomes local practice and evaluation standards. This is not merely the translation of a policy preference into rules; it is the movement of a Party-led theoretical commitment through the differentiated levels of the dual constitutional order.
The Implementation Opinions occupy a specific stage in this cascade. They translate the constitutional mandate of socialist modernization and the 3rd Plenum’s reform agenda into a governance instrument addressed to intelligent agents. Their provisions on standards, registration, testing, certification, platform rules, risk monitoring, credit evaluation, and dynamic assessment are the legal-administrative forms through which the abstract modernizing project becomes operational.
I therefore expect the Implementation Opinions to generate further specifications below the level of the document itself: technical and sectoral standards, intelligent-agent registration systems, compliance and recall mechanisms, platform duties, third-party assessment practices, and credit-evaluation criteria. The cascade will be complete only when these instruments structure local implementation and furnish the standards by which officials, enterprises, platforms, and agents are evaluated.
Limits of This Analysis
Several limitations remain. First, the Modernization paper operates at the level of semiotic structure, not empirical verification of policy outcomes. I can show how the Implementation Opinions fit within and reproduce the modernizing grammar; I cannot infer from their textual coherence that the standards, registration systems, compliance mechanisms, or applications will work as intended. The discourse-reality gap is therefore central: a document may perform modernization’s architectonic function while implementation remains fragmentary, contradictory, incomplete, or contested. Second, I must guard against the “semiotic fetish.” The system’s success in meeting its own modernization criteria cannot be treated as independent evidence of modernization’s value, because the criteria themselves produce the evidence that appears to validate them. High-quality development, new quality productive forces, and institutional adequacy are not neutral external measures in this analysis; they are categories within the modernizing lifeworld. Third, the point of the present reframing is not to ask whether the Implementation Opinions are internally coherent within that lifeworld—they manifestly are—but what the lifeworld excludes, renders invisible, normalizes, or forecloses. My work is also self-consciously resistant to the point-by-point mapping attempted here; I recognize that such a mapping may itself reproduce the “flattening” I identify in conventional analysis.[24] My comparative lecture series concludes that “AI governance is really governance of power moving through technology, and that no single system fully solves the problem”[25]—a reminder that the Implementation Opinions are as interesting for what they cannot resolve as for what they attempt. Finally, the specific Chinese AI provisions I have analyzed—the algorithmic recommendation rules, the deep synthesis provisions, the generative AI interim measures, the NFRA financial sector guidance, and the MPS cyberspace security measures—form a regulatory ecology to which the Implementation Opinions are a new addition. How this addition interacts with the existing layers, and whether it introduces tensions or redundancies, remains to be tested in implementation.
[1]Larry Catá Backer, ‘Lecture 8—Putting It All Together: Trends, Trend Lines, and Regulatory Dialectics in Comparative AI Governance’ (Law at the End of the Day, 23 June 2026) <https://lcbackerblog.blogspot.com/2026/06/lecture-8putting-it-all-together-trends.html> accessed 20 September 2026.
[2]ibid.
[3]Larry Catá Backer, ‘Lecture 5—The “Guiding State”; The Chinese Approach’ (Law at the End of the Day, 19 June 2026) <https://lcbackerblog.blogspot.com/2026/06/lecture-5the-guiding-state-chinese.html> accessed 20 September 2026.
[4]Backer, ‘Lecture 8’ (n 1).
[5]Backer, ‘Lecture 5’ (n 3).
[6]ibid.
[7]ibid.
[8]ibid.
[9]ibid.
[10]ibid.
[11]ibid.
[12]ibid.
[13]Larry Catá Backer, ‘And an Algorithm to Bind Them All? Social Credit, Data Driven Governance, and the Emergence of an Operating System for Global Normative Orders’ (Entangled Legalities Workshop, Graduate Institute of Geneva, 21 May 2018) <https://ssrn.com/abstract=3182889>.
[14]Larry Catá Backer, ‘The Semiotics of a Systems Theoretical Application of Chinese New Era Marxist-Leninism in AI Systems for the World—Reflections on Xi Jinping’s Keynote at 2026 World AI Conference’ (Law at the End of the Day, 21 July 2026) <https://lcbackerblog.blogspot.com/2026/07/the-semiotics-of-systems-theoretical.html> accessed 20 September 2026.
[15]Larry Catá Backer, ‘Reflections on Measures on Cyberspace Security Supervision and Inspection by Public Security Organs (MPS Order No 176)’ (Law at the End of the Day, 16 August 2026) <https://lcbackerblog.blogspot.com/2026/08/measures-on-cyberspace-security.html> accessed 20 September 2026.
[16]Larry Catá Backer, ‘国家金融监督管理总局发布《关于银行业保险业人工智能安全开发应用的指导意见》 [National Financial Regulatory Administration Issues the “Guiding Opinions on the Secure Development and Application of Artificial Intelligence in the Banking and Insurance Industries”]’ (Law at the End of the Day, 21 June 2026) <https://lcbackerblog.blogspot.com/2026/06/national-financial-regulatory.html> accessed 20 September 2026.
[17]Larry Catá Backer, ‘Reflections on 薛澜 人工智能技术的社会应用——治理挑战 [Xue Lan, The Social Application of Artificial Intelligence: Governance Challenges]’ (Law at the End of the Day, 22 August 2026) <https://lcbackerblog.blogspot.com/2026/08/reflections-on-xue-lan-social.html> accessed 20 September 2026.
[18]Backer, ‘NFRA Guiding Opinions’ (n 16).
[19]Larry Catá Backer, ‘Smart Regulation, Smart Society, Smart Courts, and Smart Party: The Ideology of Chinese Social Credit and its Dialectics’ (European China Law Studies Association Annual Conference 2025, University College Cork, 20 September 2025) <https://www.backerinlaw.com/Site/wp-content/uploads/2025/09/Backer_SmartRegulation_ECLSA2025_v2.pdf>.
[20]Backer, ‘Lecture 8’ (n 1).
[21]Backer, ‘MPS Order No 176’ (n 15).
[22]Backer, ‘WAIC Reflections’ (n 14).
[23]Backer, ‘Xue Lan Reflections’ (n 17).
[24]Backer, ‘WAIC Reflections’ (n 14).
[25]Backer, ‘Lecture 8’ (n 1).
Bibliography
Backer LC, ‘And an Algorithm to Bind Them All? Social Credit, Data Driven Governance, and the Emergence of an Operating System for Global Normative Orders’ (Entangled Legalities Workshop, Graduate Institute of Geneva, 21 May 2018) <https://ssrn.com/abstract=3182889>
Backer LC, ‘Lecture 1—From Algorithms to Foundation Models: What Contemporary AI is “Made of”’ (Law at the End of the Day, 13 June 2026) <https://lcbackerblog.blogspot.com/2026/06/lecture-1from-algorithms-to-foundation.html> accessed 20 September 2026
Backer LC, ‘Lecture 5—The “Guiding State”; The Chinese Approach’ (Law at the End of the Day, 19 June 2026) <https://lcbackerblog.blogspot.com/2026/06/lecture-5the-guiding-state-chinese.html> accessed 20 September 2026
Backer LC, ‘Lecture 8—Putting It All Together: Trends, Trend Lines, and Regulatory Dialectics in Comparative AI Governance’ (Law at the End of the Day, 23 June 2026) <https://lcbackerblog.blogspot.com/2026/06/lecture-8putting-it-all-together-trends.html> accessed 20 September 2026
Backer LC, ‘Musings in the Form of a Policy Brief: Bots and Automated Systems in International Institutions’ (Law at the End of the Day, 3 August 2026) <https://lcbackerblog.blogspot.com/2026/08/musings-in-form-of-policy-brief-bots.html> accessed 20 September 2026
Backer LC, ‘国家金融监督管理总局发布《关于银行业保险业人工智能安全开发应用的指导意见》 [National Financial Regulatory Administration Issues the “Guiding Opinions on the Secure Development and Application of Artificial Intelligence in the Banking and Insurance Industries”]’ (Law at the End of the Day, 21 June 2026) <https://lcbackerblog.blogspot.com/2026/06/national-financial-regulatory.html> accessed 20 September 2026
Backer LC, ‘Reflections on Measures on Cyberspace Security Supervision and Inspection by Public Security Organs (MPS Order No 176)’ (Law at the End of the Day, 16 August 2026) <https://lcbackerblog.blogspot.com/2026/08/measures-on-cyberspace-security.html> accessed 20 September 2026
Backer LC, ‘Reflections on 薛澜 人工智能技术的社会应用——治理挑战 [Xue Lan, The Social Application of Artificial Intelligence: Governance Challenges]’ (Law at the End of the Day, 22 August 2026) <https://lcbackerblog.blogspot.com/2026/08/reflections-on-xue-lan-social.html> accessed 20 September 2026
Backer LC, ‘Smart Regulation, Smart Society, Smart Courts, and Smart Party: The Ideology of Chinese Social Credit and its Dialectics’ (European China Law Studies Association Annual Conference 2025, University College Cork, 20 September 2025) <https://www.backerinlaw.com/Site/wp-content/uploads/2025/09/Backer_SmartRegulation_ECLSA2025_v2.pdf>
Backer LC, ‘The Semiotics of a Systems Theoretical Application of Chinese New Era Marxist-Leninism in AI Systems for the World—Reflections on Xi Jinping’s Keynote at 2026 World AI Conference’ (Law at the End of the Day, 21 July 2026) <https://lcbackerblog.blogspot.com/2026/07/the-semiotics-of-systems-theoretical.html> accessed 20 September 2026
Backer LC, ‘《人工智能示范法4.0》中英双语版本发布 «Bilingual Chinese-English Version of the “Model Law on Artificial Intelligence 4.0” Released»’ (Law at the End of the Day, 1 August 2026) <https://lcbackerblog.blogspot.com/2026/08/40-bilingual-chinese-english-version-of.html> accessed 20 September 2026
[1]Larry Catá Backer, ‘Lecture 8—Putting It All Together: Trends, Trend Lines, and Regulatory Dialectics in Comparative AI Governance’ (Law at the End of the Day, 23 June 2026) <https://lcbackerblog.blogspot.com/2026/06/lecture-8putting-it-all-together-trends.html> accessed 20 September 2026.
[3]Larry Catá Backer, ‘Lecture 5—The “Guiding State”; The Chinese Approach’ (Law at the End of the Day, 19 June 2026) <https://lcbackerblog.blogspot.com/2026/06/lecture-5the-guiding-state-chinese.html> accessed 20 September 2026.
[13]Larry Catá Backer, ‘And an Algorithm to Bind Them All? Social Credit, Data Driven Governance, and the Emergence of an Operating System for Global Normative Orders’ (Entangled Legalities Workshop, Graduate Institute of Geneva, 21 May 2018) <https://ssrn.com/abstract=3182889>.
[14]Larry Catá Backer, ‘The Semiotics of a Systems Theoretical Application of Chinese New Era Marxist-Leninism in AI Systems for the World—Reflections on Xi Jinping’s Keynote at 2026 World AI Conference’ (Law at the End of the Day, 21 July 2026) <https://lcbackerblog.blogspot.com/2026/07/the-semiotics-of-systems-theoretical.html> accessed 20 September 2026.
[15]Larry Catá Backer, ‘Reflections on Measures on Cyberspace Security Supervision and Inspection by Public Security Organs (MPS Order No 176)’ (Law at the End of the Day, 16 August 2026) <https://lcbackerblog.blogspot.com/2026/08/measures-on-cyberspace-security.html> accessed 20 September 2026.
[16]Larry Catá Backer, ‘国家金融监督管理总局发布《关于银行业保险业人工智能安全开发应用的指导意见》 [National Financial Regulatory Administration Issues the “Guiding Opinions on the Secure Development and Application of Artificial Intelligence in the Banking and Insurance Industries”]’ (Law at the End of the Day, 21 June 2026) <https://lcbackerblog.blogspot.com/2026/06/national-financial-regulatory.html> accessed 20 September 2026.
[17]Larry Catá Backer, ‘Reflections on 薛澜 人工智能技术的社会应用——治理挑战 [Xue Lan, The Social Application of Artificial Intelligence: Governance Challenges]’ (Law at the End of the Day, 22 August 2026) <https://lcbackerblog.blogspot.com/2026/08/reflections-on-xue-lan-social.html> accessed 20 September 2026.
[19]Larry Catá Backer, ‘Smart Regulation, Smart Society, Smart Courts, and Smart Party: The Ideology of Chinese Social Credit and its Dialectics’ (European China Law Studies Association Annual Conference 2025, University College Cork, 20 September 2025) <https://www.backerinlaw.com/Site/wp-content/uploads/2025/09/Backer_SmartRegulation_ECLSA2025_v2.pdf>.
[25]Backer, ‘Lecture 8’ (n 1).
智能体规范应用与创新发展实施意见
智能体是具备自主感知、记忆、决策、交互与执行能力的智能系统，是人工智能产品及服务的重要形态。随着大模型等新一代人工智能技术迅猛发展，智能体正加速与网络空间、物理世界深度融合，深刻改变人类生产生活方式和社会治理模式。为落实国务院《关于深入实施“人工智能+”行动的意见》，促进智能体规范应用与创新发展，制定本实施意见。
一、基本原则
以推动科技创新、提升治理能力、构建产业生态、增进民生福祉为导向，坚持安全可控，将智能体安全、可靠、可信作为发展的底线要求，贯穿智能体技术研发、应用部署与推广的全过程，切实防范系统性风险。坚持规范有序，适应智能体技术演进规律，构建与现有政策法规衔接顺畅、行业自律自治、底线红线清晰的治理体系，有序推进智能体落地应用。坚持创新驱动，加强理论创新、技术创新、工程创新联动，体系化突破智能体关键技术，完善政产学研用协同机制，构建开放共享的智能体生态，提升产业创新活力。坚持应用牵引，重点围绕科学研究、产业发展、提振消费、民生福祉、社会治理等实际需求，发挥典型应用场景示范效应，先易后难、循序渐进，促进智能体技术验证、产品迭代、应用落地。
二、夯实发展基础
夯实技术底座，健全标准体系，降低智能体研发、适配、应用门槛，为丰富智能体产品及服务奠定基础。
（一）完善技术底座
1.强化基础技术研发。持续提升通用基础模型性能，支持行业发展细分领域专用模型，形成适应不同场景和设备的模型产品矩阵。面向智能体训练与运行，提升高质量数据集供给能力。加强智能体任务理解、任务规划、工具使用、长期记忆、互认互通、群体协同等技术攻关，提升智能体泛化能力。
2.完善智能体工具链。开展智能体底层框架研究，加快研发感知、记忆、决策、交互、执行等关键组件，完善智能体研发、测试、部署、运维等工具链。发展对抗样本检测、行为异常检测等安全与治理工具，提升对智能体非合规行为的发现、干预、阻断、恢复能力。
（二）构建标准协议
3.建立智能体标准体系。制定智能体标准化工作指导文件，形成智能体标准框架，系统布局关键技术、重要产品、数据交换、应用场景、质量评测、安全保障、可信认证等标准体系，加快制定智能体与软件工具、应用服务、硬件外设接口等基础标准。加强智能体互联协议（AIP）等智能体互联关键国家标准、行业标准的推广应用。支持医疗、交通、媒体、公共安全等领域制定强制性标准。鼓励企业按照相关标准研发产品服务，提升智能体规范性。积极参与国际标准制定。
4.布局发展智能互联网。研究建立智能互联网体系架构，探索建立智能体注册平台，提供智能体数字身份管理、检索发现、能力声明等服务，支持开发者、部署方式、接口协议、合规认证等信息查询和管理。提升多智能体协同能力，研究智能体身份标识、可信互联、合规支付、安全防护、冲突解决等基础技术。发挥互联网协议第六版（IPv6）技术优势，提升智能体端到端通信能力。探索建立智能互联网监测指标体系。
三、守牢安全底线
坚持以人为本、智能向善、多元共治、安全稳妥，营造规范发展、鼓励创新的制度环境，促进智能体健康有序发展。
（一）明确产品准则
5.完善政策法规和伦理规范。加快研究智能体相关政策法规及伦理规范，发挥专业机构内容资源和审核把关优势，确保智能体行为符合法律法规及主流价值观。防止智能体利用数据优势、人格化技术实施传播不良价值观、算法压榨等行为，防范未成年人、老年人沉迷成瘾、情感依赖等风险。做好与人工智能伦理审查等制度衔接。
6.明确决策权限。在遵守法律法规、尊重社会公德和伦理规范前提下，厘清仅限用户本人决策、需由用户授权决策和智能体自主决策等各种决策方式的合理边界及所需权限。确保用户对智能体自主决策享有知情权和最终决策权，智能体执行操作不得超出用户授权范围。
7.加强行为管控。发展规则内嵌、行为围栏等技术，确保智能体在公共场所、隐私场所、专门场所等的行为合法合规。探索利用区块链等技术，建立重要应用场景智能体行为可验证、可追溯机制，防范智能体不当行为引发重大风险。
（二）防范安全风险
8.提升内生安全能力。研究智能体数据安全、个人信息保护、密码防护、攻击检测、权限管理、行为控制等安全技术，提升智能体系统安全保障能力，防范数据投毒、隐私泄露、算法篡改、系统漏洞、运行失控等安全风险。研究智能体安全检测技术，探索建立智能体安全评估体系。
9.加强供应链安全。制定智能体开发、部署、应用、维护等全周期安全规范，加强模型接入、应用程序接口调用、扩展工具使用等环节安全管理。探索建立智能体供应链安全信息共享和预警机制，及时发布风险提示，提升安全保障能力。
10.化解应用衍生风险。完善智能体常态化风险识别、预警及干预机制，强化人机协同审核、拦截阻断等风险处置能力，防范系统性安全风险。强化智能体应用安全管理，避免智能体被用于自动化攻击、隐私侵犯、虚假信息生成传播、网络诈骗等违法犯罪行为。
（三）完善治理体系
11.构建分类分级治理框架。根据应用场景和潜在影响，审慎稳妥开展智能体分级治理。对于敏感领域及重点行业，由网信部门联合行业主管部门确定开放场景，根据相关法律法规、监管要求和安全防护标准，实行备案、检测、问题产品召回等管理措施。对于部分生活娱乐、日常办公等低风险领域，完善智能体评估测试工具，通过合规自测、信息报告、分发平台管理、行业自律等实现高效治理。
12.健全合规服务体系。强化智能体风险监测预警、检测评估、咨询、认证等专业服务供给，引导行业积极研发智能体监测工具。开展智能体功能、性能、质量、合规等第三方评测服务，推动认证与检测结果互通互认，为用户选择智能体提供参考。编制并发布智能体技术及应用成熟度报告，为产业研发应用提供参考。
（四）强化行业自律
13.引导行业加强自律。鼓励行业组织、主要企业联合制定行业自律规则，明确智能体功能合规、算法治理、知识产权保护、公平竞争等规范细则。指导智能体开发平台、分发平台、服务提供者建立公平合理的平台规则、用户服务协议及隐私政策，明确供需双方权责，保障产业健康发展。加强智能体应用风险宣传教育，提升用户安全意识。
14.探索信用评价机制。指导行业组织建立智能体市场主体自愿参与的信用评价机制，对于技术滥用、诱导消费、虚假宣传、隐瞒缺陷信息等行为进行信用评价，依法依规开展失信惩戒。引导智能体开发者、开发平台、分发平台、服务提供者等参与信用评价，共同营造良好发展环境。
四、强化应用牵引
积极稳妥推动智能体典型场景应用，牵引技术产品优化提升，探索形成可复制、可推广的智能体落地应用模式。
（一）科学研究
15.科研探索。研发理论推演、模拟仿真等智能体，挖掘潜在技术路径。强化智能体信息关联整合、知识体系构建等能力，提升自然科学、哲学社会科学研究发现能力。促进智能体与科学仪器、实验平台融合，实现方案设计、实验操作、数据处理、结果分析等全流程智能化。
16.研发辅助。发展软件开发智能体，提升需求分析、架构设计、代码生成与测试等全流程开发能力。促进智能体与计算机辅助设计（CAD）、计算机辅助工程（CAE）等软件结合，提供设计方案生成、仿真验证、参数调优等功能。
（二）产业发展
17.智能制造。研发生产管理智能体，动态优化生产排程、资源分配和工序衔接，推动智能体在工业互联网领域应用，提升企业精细化管理水平。提升智能体工艺参数优化、加工精度检测、产品缺陷识别等能力，促进智能体与数控机床、工业机器人、自动化产线等融合，促进提质增效降本。
18.能源资源。研发大气、水体、土壤、噪声等环境要素感知智能体，提升自然灾害、环境污染等风险预警能力。依托智能体强化对国土空间资源全周期管理能力。依托智能体实现能源、金属等矿产资源高效勘探。发展电力调度、用电监测、电网维护等智能体，提升电力资源使用效率。
19.交通运输。研发交通安全监管、应急指挥调度等智能体，提升违章违规行为识别、交通基础设施风险预警、重点车辆（船舶）监管、事故快速响应等能力。优化交通监测调度智能体性能，发展交通载运工具管控智能体，提升路网、水网、空域的通行效率。
20.农业生产。研发农业服务智能体，开展农技指导、病虫害诊断与防治等服务。推动智能体在种植养殖、高效育种等环节应用，推进农业智能化转型。推动智能体与智能农机具、智慧大棚、农业服务平台融合，提升农业生产效率。
21.金融服务。研发金融风控智能体，提升信贷审批、交易监控、账户安全等环节风险识别能力。完善智能体异常检测、合规审计功能，提升信贷违约预测、信用卡盗刷拦截、反洗钱监测等能力。
（三）提振消费
22.终端应用。推动智能体赋能互联网应用及服务，优化在线购物、出行导航、生活缴费、日常办公等服务体验。推动智能体与手机、电脑、汽车、家居、可穿戴、消费级机器人等终端设备协同发展，提升跨应用、跨设备任务完成能力。
23.文化旅游。研发文学、音乐、绘画、视听、演艺等内容创作智能体，促进优秀文化传播推广。发展智能导览、多语种翻译、适老适残等旅游服务智能体，提升旅游服务水平。
24.商业服务。提升智能体客服能力，提供7×24小时咨询、预约、售后等服务。发展导引、清洁、仓储、配售等具身智能体，提升餐饮、零售、住宿、物流等商业场所的运营效率。探索通过具身智能体提供低成本家政、养老、托育、助残等服务。
（四）民生福祉
25.教育教学。探索课件生成、作业批改、学情分析等智能体，提高教师工作效率。依托智能体开展个性化学习方案制定，完善智能导学、答疑辅导、虚拟助教等功能。支持在线教育平台研发智能体，提供终身学习服务。
26.医疗健康。提升医学影像分析、疾病诊断推理、定制化诊疗方案生成等医疗辅助智能体性能，探索药品管理、手术排程、病历管理等智能体，提升医疗服务效率。稳步发展预问诊、报告解析等智能体，提升患者体验。
27.人力资源。探索智能体在就业促进、技术技能人才培养评价、劳动关系公共服务等领域应用，提升就业服务能力。发展社会保险、劳动争议仲裁、欠薪治理等智能体，保障劳动者合法权益。
28.信息服务。探索智能体在网络内容建设管理中的应用，鼓励信息发布部门和内容传播平台研发用户分析、选题策划、采编加工、分发推荐、智能审核、舆论引导、情绪疏导、实时翻译等智能体，实现多模态信息、跨领域信息的高效整合。
（五）社会治理
29.政务服务。探索事项辅助审批智能体，推动政务审批流程智能化。发展政策咨询智能体，提供全天在线的政务咨询、流程指引等服务。探索主动推送适配政策、服务提醒及办理指南，加快从“人找服务”向“服务找人”转型。
30.司法服务。探索全流程办案辅助智能体，提升案件材料梳理、案件信息录入、证据审查、辅助法律文书生成等能力。发展法律宣传、法律咨询、法律监督等智能体，为群众提供高效便捷的在线司法服务。
31.公共安全。探索监测预警、应急处置、救援调度、协同治理等智能体，提升安全生产监管和防灾减灾救灾等能力。提升智能体异常行为识别、潜在威胁预警、动态防控处理能力，维护公共安全。推动具身智能体在灾害救援、安防巡检、危险品处置等领域落地应用。
32.城市治理。探索智能体在城市规划、建设与治理环节应用，支撑智能建造、房屋管理、城市基础设施安全运行等工作，提升城市治理专业化水平，提升城市人居环境质量。
33.招标投标。探索招标投标智能体，实现招标投标活动全链路智慧管理，保障全过程规范高效。提升招标投标交易、服务和监管的智慧化水平，实现招标科学合理、评标公平公正和监管穿透高效。
五、建设创新生态
畅通供需渠道，促进研发侧、需求侧高水平互动，形成市场牵引、内驱发展的智能体产业生态。
（一）促进产业合作
34.培育开源创新力量。引导国内人工智能开源社区加强智能体布局，开展智能体与开源芯片、开源操作系统、开源大模型兼容适配。引导企业、高校、科研机构积极参与智能体框架、交互接口、工具链等开源项目，推动技术体系融通发展，加快提升国际影响力。
35.搭建产业协作平台。发挥智能体相关生态联盟、技术验证实验室等产业协作平台的作用，协同产业链上下游开展智能体共性技术研发、标准制定、评估认证等工作，开展智能体技术与产业应用复合人才培养。引导互联网应用、智能终端等领域企业共建生态，探索建立互利共赢的合作模式。
（二）强化应用推广
36.构建应用推广渠道。推动建立智能体软件商店、行业供需信息发布平台，引导智能体企业积极发布产品，形成集聚效应。开展智能体应用供需对接活动，采取公开招标、揭榜挂帅等方式吸引智能体企业定制化开发相应产品。引导整机、软件等企业基于智能体研发产品和服务，培育用户市场。
37.推进重点场景开放。推动重点领域开放智能体应用场景，在产业集聚区、重点行业、重点领域开展智能体应用试点，打造一批具有引领作用的示范项目。发展市场化、专业化的智能体技术转化服务机构，探索智能体应用场景，提升技术成果转化效率。促进行业数据共享开放，支撑重点场景智能体训练部署。
38.积极培育全球生态。依托世界人工智能大会、世界互联网大会等国际平台，交流展示智能体技术创新成果。推动终端设备、软件企业适配智能体，引导相关企业做好海外合规建设，推动智能体适应当地法律法规和文化习俗。
六、保障措施
国家网信办、国家发展改革委、工业和信息化部会同有关方面加强统筹谋划，强化资源整合和力量协同，完善配套政策，形成工作合力，推动重点任务落实落地。建立并完善智能体发展评价指标体系，加强智能体规范应用与创新发展的监测评估、滚动实施和动态调整。
No comments:
Post a Comment