Beijing has moved AI agents from launch-event vocabulary into a government document.
On July 23, the Beijing Municipal People's Government published Several Measures of Beijing Municipality on Accelerating Agent-Led Development. The document was jointly issued by the municipal development and reform commission, the cyberspace office, the science and technology commission and the economic and information technology bureau. Its reference number is Jing Fa Gai [2026] No. 1185, and it lays out ten measures.
The important signal is not a generic pledge to support AI. Beijing has placed terms usually heard in developer and startup circles directly into policy language: Harness Engineering, OPC (one-person company), token economy, TaaS, AaaS, RaaS, FDE and AIP. The policy target has moved below model vendors, toward the middleware, toolchains and small teams that make agents usable.
Filling the Middleware Layer
The document opens by saying Beijing wants to promote agent innovation, cultivate new forms of intelligent economy and accelerate its role as a global AI innovation hub. At the technical level, the first measure points to foundation-model capabilities such as online learning, continual learning, autonomous evolution, long-horizon tasks, complex reasoning and planning.
The second measure is more concrete. It says Beijing will support innovators in implementing Harness Engineering, strengthening the shared base for agents through context engineering, task persistence, multi-agent collaboration and system scalability.
In plain language, Beijing is looking beyond model parameters and benchmark rankings. It wants to build the glue layer that lets agents run long business tasks: memory, context, tool calls, task queues, cross-model and cross-chip adaptation, safety controls and logs. Many corporate agent trials fail at precisely that layer. A model that can answer questions is not automatically a system that can complete a workflow.
The measures also call for open and controllable interconnection protocols, development frameworks, basic toolchains and key capability components, while naming standards such as the Agent Interconnection Protocol, or AIP. If this layer matures, the beneficiaries will include not only model companies but also workflow vendors, MCP connector builders, enterprise integration teams and security-sandbox providers.
OPCs Move Into the Main Program
Measures three through five combine application scenarios with entrepreneurship. Beijing calls for scientific AI assistants, “AI scientists,” autonomous laboratories and AI OS, and it backs agent deployment in EDA, quality inspection for flexible-display production lines and electronics supply-chain scheduling.
The more Beijing-specific item is the OPC. The document supports new entrepreneurial models represented by OPCs, or one-person companies, including OPC communities, full-cycle service stations, streamlined registration, one-click form filling and automatically generated standardized articles of association.
This is not an isolated move. The Beijing News reported on July 22 that Beijing had already released an OPC action plan. Within the year, the city aims to cultivate at least 10 professional OPC growth communities and serve more than 500 high-quality OPC innovators. High-performing communities can receive rewards of up to RMB 2 million, while resident companies can receive three consecutive months of services worth up to RMB 100,000, covering token use, intelligent computing power and data procurement.
In the agent economy, the logic is straightforward. If one person can use AI to call design, R&D, sales, customer-service and finance tools, the city wants to connect such teams to office space, compute, compliance, financing and orders. Incubators used to serve companies. Their client is starting to become one person plus a set of agents.
The Token Economy Is Not Crypto Talk
The sixth measure refers to the token economy, a phrase that can be misread as virtual assets. In this document, tokens are closer to a unit for large-model usage, task throughput and intelligent-service quality.
Beijing says it will promote models such as Token as a Service, Agent as a Service and Result as a Service, while exploring token service-quality evaluation and billing standards. It also encourages a shift from pricing by token consumption toward value-based pricing. Background reporting from The Beijing News adds several industry numbers: in the first half of the year, Beijing's digital economy value added grew 7.8%, the core digital-economy industries grew 9.8%, new computing capacity reached 22,000 P, and the city's first “token factory” averaged more than 200 trillion token calls per day.
That shift is practical. For companies buying AI services, the price per million tokens is no longer enough. If an agent can close a loop across approvals, customer service, search, graphics and code changes, buyers care more about task completion rate, latency, retry behavior and labor saved. By writing value-based billing into the policy, Beijing is acknowledging that agents may change how software is purchased.
Governance Is in the Same Frame
The seventh measure covers safety governance: classified and graded supervision of agents, regular crackdowns on malicious misuse, AI industry legislation, agent security service platforms, testing ranges, trusted sandboxes, vulnerability scanning, security testing, attack-defense exercises and certification services.
Placed beside the industry-support measures, the boundary becomes clearer. The more agents can call tools and execute tasks across systems, the more they need identity, permissions, auditability and sandboxes. Beijing is encouraging FDE co-creation, lightweight OPC entrepreneurship and token factories while also writing model-governed safety and trusted sandboxes into the base layer. Policymakers appear to recognize that agent risk comes not only from generated content, but also from the ability to act.
Three Tests to Watch
First, watch whether AIP, toolchains and security sandboxes turn into named projects. Without the middleware layer, agent policy can easily remain conceptual.
Second, watch how token vouchers and service vouchers are distributed. Subsidies based only on usage can encourage volume inflation. Subsidies tied to task delivery, industry scenarios and customer renewal would look more like intelligent-economy infrastructure.
Third, watch whether OPC communities generate real orders. One person plus AI can form a company, but survival still depends on customers, compliance, cash flow and delivery records.
The larger signal from Beijing's document is that agents are being treated as the next form of software and industrial organization. Model vendors, tool developers, startup communities, computing platforms and regulators are now described in one policy language. AI agents are no longer just the “next application” in a product manager's deck; they are starting to receive city-level budgets, standards and accountable agencies.
Sources: Beijing Municipal People's Government policy document, The Beijing News, Beijing Daily client, CocoLoop; checked issuing agencies, policy number, ten measures, OPC and token-economy support scope, first-half digital-economy and computing-capacity data.