Beijing has moved agents out of the launch-event vocabulary and into a government policy file.
On July 23, the Beijing Municipal People's Government website published Several Measures of Beijing Municipality on Accelerating Agent-Led Development. The document was issued by four city agencies, including the Development and Reform Commission and the cyberspace, science and industry authorities. It carries the policy number Jing Fa Gai [2026] No. 1185 and is organized into ten measures.
The important signal is not simply that Beijing supports AI. The policy text uses terms that until recently mostly lived in developer and startup circles: Harness Engineering, OPC, token economy, TaaS, AaaS, RaaS, FDE and AIP. That shifts the policy target from model companies alone to the agent middleware layer, toolchains and small teams building around them.
Filling the agent middle layer
The first measure points to stronger foundation-model capabilities: online learning, continual learning, autonomous evolution, long-horizon tasks, complex reasoning and planning. The second is more operational. It says Beijing will support innovators in harness engineering, including context optimization, task persistence, multi-agent collaboration and scalable systems.
In practice, that means Beijing is looking at the glue that lets agents run long jobs: memory, context, tool calls, queues, cross-model and cross-chip adaptation, safety controls and logs. Many enterprise pilots fail at exactly this layer. A model that can answer a question is not automatically a system that can finish a business workflow.
The policy also calls for open and controllable interconnection protocols, development frameworks, basic toolchains and key components, and names Agent Interconnection Protocol standards. If this layer matures, the beneficiaries will include workflow companies, MCP connectors, enterprise integration vendors and security sandbox providers, not only model labs.
OPC joins the main track
Measures three through five connect applications with entrepreneurship. Beijing wants agent systems in scientific assistants, “AI scientists,” autonomous laboratories, AI OS, EDA, flexible-display quality inspection and electronics supply-chain scheduling.
The more local feature is OPC, or one-person company. The document says Beijing will support new startup models represented by OPCs, build OPC communities and full-cycle service stations, streamline registration and offer one-click filing and standardized articles of association.
This follows an OPC action plan reported by The Beijing News on July 22. Beijing plans to cultivate at least 10 specialized OPC growth communities this year and serve more than 500 high-quality OPC innovators. Strong-performing communities can receive up to RMB 2 million, and resident companies may receive three months of free support worth up to RMB 100,000 for token use, intelligent compute and data procurement.
For the agent industry, the meaning is straightforward: if one person can use AI to call design, engineering, sales, customer-service and finance tools, the city wants to connect that team to office services, compute, compliance, financing and orders. Incubators used to serve companies; the new target can be one person plus a set of agents.
Token economics here is not crypto language
The sixth measure discusses the token economy. In this policy, token is closer to a unit for model calls, task throughput and service quality than to a virtual asset.
Beijing proposes Token as a Service, Agent as a Service and Result as a Service, and says it will explore quality assessment and billing rules for token services. The text encourages a move from billing by token consumption toward value-based charging. The background figures are also clear: Beijing's digital-economy value added grew 7.8% in the first half, core digital industries grew 9.8%, new compute capacity reached 22,000P, and the first “token factory” averaged more than 200 trillion token calls per day.
That shift matters for enterprise buying. Price per million tokens is no longer enough if an agent can close a loop across approvals, customer support, search, design or code changes. Buyers care about completion rate, latency, retries and labor saved. Beijing's value-pricing language recognizes that agents may change the way software is purchased.
Governance is in the same frame
The seventh measure covers safety governance: classified and tiered agent supervision, routine crackdowns on malicious use, AI industry legislation, agent safety-service platforms, test ranges, trusted sandboxes, vulnerability scanning, security testing, red-team exercises and certification.
Read with the industrial measures, the boundary is clear. The more agents can call tools and act across systems, the more they need identity, permission control, auditing and sandboxes. Beijing is encouraging FDE co-creation, lightweight OPC startups and token factories while also putting model-governing-model approaches and trusted sandboxes into the foundation.
Three implementation details to watch
First, watch whether AIP projects, toolchains and sandboxes become concrete. Without a middle layer, agent policy can remain a slogan.
Second, watch how token vouchers and service vouchers are distributed. Subsidies based only on usage invite volume chasing; subsidies tied to delivery, industry scenarios and renewals look more like infrastructure for an intelligent economy.
Third, watch whether OPC communities bring real orders. A person can register a company with AI, but survival still depends on customers, compliance, cash flow and delivery records.
The strongest signal in Beijing's document is that agents are being treated as the next form of software and industrial organization. Model vendors, tool developers, startup communities, compute platforms and regulators are now placed in one policy language. Agents are no longer just “the next app” in product pitches; they now have city-level budgets, standards and accountable departments.
Sources: Beijing Municipal People's Government policy file, The Beijing News, Beijing Daily app, CocoLoop; checked the issuing bodies, policy number, ten measures, OPC and token-economy support language, first-half digital-economy and compute-capacity figures.