On 11 September 2026 the Ministry of Industry and Information Technology issued the Implementation Plan for the AI Plus Software Special Action (the Implementation Plan) and held a dedicated press conference that day to explain its principal points. The Plan sets out a quantified roadmap to 2028 and 2030: promotion and application reaching 20,000 software enterprises above the designated size, 100 accumulated intelligent technical upgrading projects in software enterprises, and 100 benchmark agent-software applications. Beyond the policy benefits, the Plan also sets out several compliance requirements to which companies must respond in advance.
Policy objectives and industry background
The Implementation Plan sets two staged objectives. By 2028, the level of intelligence in the software and information technology services industry is to rise markedly, with promotion and application reaching 20,000 software enterprises above the designated size, 100 accumulated intelligent technical upgrading projects, 100 benchmark agent-software applications in key industries, and more than five key open-source projects incubated. By 2030, the integration of AI with the software and information technology services industry is to reach a new level, key software is to be comprehensively upgraded, and new business forms such as intelligent programming, agent software and intelligent services are to become new growth poles, moving the industry up the global value chain.
The timing of the policy reflects the size of the industry base. In 2025 the revenue of China’s software industry reached RMB 15.48 trillion, 6.2 times that of 2012, with a compound annual growth rate of 15.1%; and by June 2026 daily token invocations in China had exceeded 500 trillion.
Six areas of deployment
The Implementation Plan makes provision in six areas, which may be summarised as three upgrades, one foundation, one environment and one new business form: advancing the transformation of software production, optimising the chain of intelligent programming tools and organising intelligent technical upgrading in software enterprises; accelerating the intelligent upgrading of software products, including basic software and industrial software; cultivating the new business form of agent software, developing general, specialised, vertical and industrial agents; expanding intelligent software services, developing model-as-a-service and agent-as-a-service; consolidating the foundations of intelligent software development, promoting collaborative open-source innovation, strengthening the support of computing services and building high-quality datasets; and improving the development environment, including key standards, talent cultivation, employment-friendly development and security capacity.
The Plan calls for three transitions: software development from manual coding to human-machine collaboration; software products from passive response to active execution; and software services from delivery of a product to delivery of value. The vehicle for “active execution” is precisely agent software.
Three questions raised by agent software
How liability is allocated once execution becomes active. The Plan proposes to cultivate industrial agents that are “secure, reliable and verifiable in behaviour”, and to develop general agents and specialised vertical agents for office collaboration, industrial production and other scenarios. The qualifier “verifiable in behaviour” indicates that the regulatory expectation of an agent is not merely capability but traceability, verifiability and attributability. The shift from passive response to active execution means that an agent will autonomously invoke tools, operate systems and initiate transactions. Once loss occurs, it becomes difficult to determine whether the cause is a defect in the software, a user instruction or a third-party tool. This is consistent with the emphasis in the AI Safety Governance Framework 3.0 on distinguishing the responsibilities of developer, provider and user. A company building agent software must define the boundaries of liability and the exclusions in its contracts, and retain behavioural logs to satisfy the “verifiable” requirement; the liability provisions in software licences and service agreements must be rewritten for the purpose and cannot follow a conventional software template.
Review for listing on agent app stores. The Plan proposes to build an agent software app store and a library of skill packages (Skills), and to regulate review for listing and operations on the platform. This is an easily overlooked signal with far-reaching consequences: once a company operates a distribution platform for agents or skill packages, it is no longer merely a technology provider but assumes platform governance responsibilities akin to those of an app store, and must answer for the compliance, safety and traceability of the content listed. The platform operator should establish admission criteria for skill packages, a content review process, a mechanism for removal of non-compliant content and a chain of accountability. Drawing on existing regulatory practice for app stores and mini-program platforms, building the review system in advance is far less costly than remediating afterwards.
The security boundary of open-source collaboration. The Plan makes “promoting collaborative open-source innovation” an important element of consolidating the development base, and proposes to incubate more than five key open-source projects and to cultivate open-source zones for agents. Open source accelerates the ecosystem but also raises practical questions of licence compliance and security responsibility: how the type and restrictions of an open-source licence are to be observed, and how liability is to be allocated when an open-source component fails. This echoes Article 13 of the Supreme People’s Court Opinion on AI disputes, which grants an appropriate exemption to a developer that provides code modules free of charge on an open-source basis and publicly explains their functions and safety risks. Open source is not immunity: disclosure and compliance documentation are the precondition of the exemption. A company participating in an open-source project should review the type and restrictions of the licences it uses and establish an open-source compliance review process; as a publisher, it should complete its safety risk disclosure documentation at the same time.
The employment question addressed directly
In response to the widely debated question of AI replacing programmers, the Plan expressly proposes employment-friendly development of the software industry: maintaining an employment-friendly orientation, guiding software enterprises to optimise production organisation around human-machine collaboration, implementing job redesign and skills training in step, and actively developing new roles. For a software company, employment-friendly development is both a policy orientation and a practical question of labour relations management: job adjustment, skills training, variation of employment contracts and lawful routes to workforce reduction must all be designed in advance, to avoid the risk of labour disputes in large-scale restructuring.
How companies should prepare
The Implementation Plan is an industrial policy and special action plan, advanced principally through guidance, support and standards development, and it does not itself impose administrative penalties. Its requirements on standards, platform review and security, however, will gradually be translated into specific regulatory and standards obligations. Companies building ordinary SaaS rather than agent software should also take note: the Plan’s objective is the comprehensive intelligent upgrading of key software, and AI functions embedded in a product — intelligent programming, intelligent customer service, intelligent analysis — equally engage data compliance, content safety and the allocation of liability.
In practical terms, preparation may proceed along several lines: rewrite software licences and service agreements to define liability and exclusions in agent scenarios and to retain behavioural logs; where operating a distribution platform for skill packages or agents, establish admission, review, removal and traceability mechanisms; review the type and restrictions of open-source licences and, as a publisher, complete disclosure of function and safety risk; monitor the development of key standards on intelligent programming, agent interfaces and intelligent services and align in advance; when building high-quality datasets, implement the requirements on lawful provenance and personal information protection; and design job redesign, skills training and labour relations adjustment together, to forestall labour disputes. Building safety, verifiability and open-source compliance into the product both satisfies the policy orientation and improves the prospects of inclusion in benchmark applications and pilot programmes.
For assistance with liability and licensing design for agent software, platform listing review systems, open-source compliance, data compliance, or corporate governance and equity structure and data compliance for software companies, please contact Geason Law Offices. Further reading: the commentary on the AI Safety Governance Framework 3.0 and the commentary on the Supreme People’s Court opinion on AI disputes.
This article is prepared on the basis of the Implementation Plan for the AI Plus Software Special Action issued by the Ministry of Industry and Information Technology on 11 September 2026 and the related press conference. It is for reference only and does not constitute legal advice. Individual matters should be assessed on their own facts and evidence.