China's MIIT orders AI into 20,000 software firms by 2028
Beijing's software regulator just turned the industry's shift from copilots to agents into policy — with quotas, deadlines, and a checklist.
China's Ministry of Industry and Information Technology has issued a dedicated "AI + Software" action plan that sets hard adoption targets for putting AI into the country's software industry. By 2028, the plan calls for AI-powered development to reach 20,000 above-scale software companies, 100 corporate smart-transformation projects, and 100 flagship agent-software applications in key industries — alongside at least 15 hardware-software adaptation centers and at least five quality open-source projects incubated. By 2030, the targets graduate from adoption to completion: key software is expected to finish its AI upgrade entirely, with agentic coding tools, agent software, and intelligent services positioned as the industry's new growth engines.
The plan's two main levers say a lot about where Beijing thinks software is going. The first is agentic coding: tools are explicitly directed toward project-level, full-lifecycle autonomous development rather than autocomplete, deep integration with code hosting, cloud services, and open-source communities, plus consumer-friendly coding tools that lower the barrier for non-developers. The second is agent software as a product category in its own right — engineering work on runtime frameworks, reliable execution of complex tasks, and multi-agent coordination, distributed through agent app stores and skills libraries, across three tracks: general-purpose agents for office and daily services, on-device agents with lightweight reasoning and privacy protection, and industrial agents for core manufacturing tasks that the plan says must be safe, reliable, and behavior-verifiable.
The rest of the document fills in the scaffolding. Operating systems and databases get intelligence upgrades plus joint optimization with chip and server vendors; industrial software — CAD, CAE, EDA — is slated for deep AI fusion; and security is written into the development lifecycle itself, with large models tasked with vulnerability discovery, automatic defect detection, and risk grading of existing software. Standards work rounds it out: maturity grading for coding tools, agent interface standards for hardware-software coordination, and specifications for model-as-a-service and agent-as-a-service offerings.
Our take: this is the state sequencing the same transition the rest of the industry is living through, but with delivery quotas attached. Quotas like these do two things — they pull vendor marketing ("agentic") straight into procurement checklists, and they steer funding toward whatever a ministry certifies as a flagship. The genuine win could be the industrial-agent requirement: "behavior-verifiable" is a real engineering bar, and if it becomes a de facto certification standard, it will shape agent tooling well beyond China. We tracked the earlier round of this playbook in August — China's MIIT sets hard targets for an AI service-provider army — and this plan extends it from service providers to the software industry itself.
What to watch: which provinces publish matching subsidy programs first, and whether the behavior-verifiable industrial-agent requirement hardens into a national certification standard.
Would hard state quotas make agents ship faster — or just make everything get labeled an agent? Tell us in the comments.
