Beijing AI Agent Policy Makes Harness Engineering, Token Economics, and OPCs Official

Beijing has introduced a ten-point AI agent policy that reportedly puts Harness Engineering, Token economics, and one-person companies into an official framework for the first time.

The Beijing AI agent policy matters because it treats agents as an economic system, not another category of software. Its focus extends beyond models to execution environments, service contracts, payment methods, specialized businesses, and devices.

Reporting published on July 23 identifies the document as Measures to Accelerate Development Led by AI Agents. However, a complete copy was not yet visible in the searchable policy databases reviewed for this article. The most specific provisions therefore require cautious treatment until Beijing publishes a canonical text.

The reported direction still fits Beijing’s existing policy sequence. Municipal authorities already support AI-driven one-person companies, industrial agents, computing credits, data services, and agent deployment in manufacturing.

China’s national government has also set a broader target. Its AI Plus policy calls for adoption of new intelligent devices and agents to exceed 70 percent by 2027.

The tension is clear. Cloud AI businesses usually meter consumption through tokens, while Beijing reportedly wants suppliers to charge for completed value. That shift sounds attractive, but it transfers commercial and technical risk from customers to agent providers.

This is not simply a contest between Beijing and another city. The central conflict is between consumption-based AI and outcome-based AI. The new policy attempts to move the market from selling model activity toward selling completed work.

That creates pressure for model providers, agent startups, enterprise buyers, and developers. Each group must decide how to define a successful outcome, prove that an agent produced it, and assign responsibility when automated work fails.

What the Beijing AI Agent Policy Actually Changes

The policy’s most important move is to treat the engineering around a model as a first-class industrial layer.

According to the initial report, Beijing’s measures contain ten provisions and explicitly reference Agentic AI, Harness Engineering, AI operating systems, field deployment engineers, OPCs, and several service models.

Harness Engineering means designing the systems that make a model useful during sustained work. Those systems include context management, tools, permissions, memory, evaluations, feedback loops, and execution environments.

That language reflects a change already underway among leading AI laboratories. OpenAI described harness engineering as building environments, specifying intent, and creating feedback loops that let coding agents perform reliable work.

In OpenAI’s reported internal experiment, agents produced every line of code for a software product. Human engineers concentrated on specifications, structure, feedback systems, and operational controls.

Anthropic has reached a related conclusion. Its work on long-running agents found that capable models still struggle across separate context windows without structured handoffs and persistent artifacts.

Beijing’s reported policy translates that engineering lesson into an industrial signal. The model alone is no longer the complete product. Reliability increasingly depends on the surrounding architecture.

This distinction matters in procurement. Two services can use similar foundation models yet deliver different results because their harnesses provide different tools, memories, controls, and verification steps.

The document reportedly uses a local translation that presents the harness as an engineering layer for directing or controlling agents. The exact English equivalent remains less important than the policy choice behind it.

Beijing is telling companies that orchestration, observability, and evaluation deserve support alongside model development. That favors teams building dependable systems around existing models, not only companies training larger models.

The policy also reportedly encourages agents inside phones, smart glasses, automobiles, and other terminals. This expands the target market beyond browser-based assistants and enterprise dashboards.

A device-based agent operates under tighter constraints than a chatbot. It must manage latency, permissions, connectivity, sensor data, battery use, privacy, and physical consequences.

Those constraints make the harness even more important. An agent controlling a car function or accessing a camera needs stricter boundaries than an assistant drafting an email.

This also explains why the policy references AI operating systems. An agent running across applications needs a stable layer for identity, permissions, tool access, memory, and communication.

The reported inclusion of an agent interoperability protocol points in the same direction. A market of specialized agents cannot scale if every service uses incompatible identities, messages, and control systems.

Still, the policy does not establish a technical standard merely by naming the category. Developers need specifications, reference implementations, conformance tests, and security rules before interoperability becomes dependable.

Beijing’s intervention should therefore be read as agenda setting. It identifies the layers officials want companies to build, but implementation will determine whether those layers become infrastructure or remain policy vocabulary.

Why Beijing Is Moving From Tokens to Outcomes

The proposed billing change challenges the simplest commercial model in generative AI: charging customers for how much computation they consume.

A token is a small unit of text processed or generated by a language model. Providers commonly use input and output token counts to meter API consumption.

That method is understandable because tokens provide a measurable proxy for computing work. It also leaves customers carrying most execution risk.

If an agent consumes a large context, calls several tools, repeats failed steps, and returns an unusable answer, the customer still pays for the activity.

The Beijing AI agent policy reportedly calls for a shift from token-consumption billing toward value-based billing. Under that approach, payment would depend more directly on completed tasks, accepted results, or measurable business outcomes.

The policy also reportedly encourages Task-as-a-Service, Agent-as-a-Service, and Reasoning-as-a-Service. These categories package different parts of automated work as services.

Task-as-a-Service would sell completion of a defined job. Agent-as-a-Service would provide ongoing access to an agent with tools and workflows. Reasoning-as-a-Service would package more complex planning or decision support.

The labels are not yet standardized commercial categories. Providers can interpret them differently, and buyers should not assume that every product using the same label offers comparable accountability.

The important change is the unit of value. A token-based supplier sells processing. An outcome-based supplier sells the result of processing.

That creates an immediate measurement problem. A coding agent can produce a pull request, but completion does not mean the code is correct, secure, maintainable, or aligned with product requirements.

A research agent can deliver a report, but word count is not a meaningful measure of accuracy. A customer-service agent can close a ticket while leaving the customer dissatisfied.

Outcome-based contracts therefore require acceptance criteria. Buyers and suppliers need to specify what counts as success, who validates it, and how exceptions affect payment.

Harness Engineering becomes the mechanism that makes such contracts possible. The harness can record actions, enforce permissions, run evaluations, store evidence, and route uncertain cases to humans.

Without those controls, value-based billing becomes a marketing promise. A provider can claim that an agent completed work without giving the customer enough evidence to assess the result.

The policy’s token-economic language should not be confused with cryptocurrency. Here, Token economics concerns the cost, allocation, and business value of model usage.

The central question is not whether an agent consumed many tokens. It is whether those tokens produced an outcome worth their cost.

That distinction can reward efficient engineering. Providers gain an incentive to route simple work to smaller models, reduce unnecessary context, reuse verified results, and stop unproductive loops.

It can also encourage hidden compromises. A vendor paid per completed task might reduce verification, choose a cheaper model, or classify partial work as successful.

Enterprise buyers will need audit rights and quality measures. They should evaluate completion rates together with error rates, review effort, latency, security incidents, and downstream rework.

Knowledge-intensive work presents an additional challenge. Its value often appears later and depends on information distributed across documents, conversations, and previous decisions.

A dependable AI knowledge base can help preserve the evidence behind an agent’s output. It does not solve attribution, but it makes review more practical.

Outcome billing will become credible only when both sides can inspect how an agent reached a result. Beijing’s policy recognizes the destination, while the market still needs to build the measurement system.

Harness Engineering Puts Agent Providers Under Pressure

Once policy recognizes the harness as part of the product, providers cannot blame every failure on the underlying model.

This is the sharpest commercial consequence of Beijing’s move. The market has often treated model capability as the main explanation for agent performance.

That explanation becomes weaker when providers choose the prompts, tools, permissions, memory systems, routing logic, and evaluations surrounding the model.

OpenAI’s reported experiment illustrates the shift. The company said agents wrote a software product containing about one million lines of code in roughly one-tenth of the expected manual development time.

Those figures describe one internal project, not an independent industry benchmark. They still show why organizations now invest in the environment around coding agents.

Anthropic’s research exposes the opposite side. Even a capable model can fail during long projects when it attempts too much work at once or loses continuity between sessions.

Its response involved initializer agents, incremental tasks, progress files, and structured handoffs. Later work added planner, generator, and evaluator roles for sustained application development.

These practices are not cosmetic product features. They determine whether an agent can resume work, detect errors, and respect constraints.

Beijing’s reported recognition of Harness Engineering pressures vendors to explain those mechanisms. Enterprise customers can ask what the agent remembers, which tools it can access, and how its output is tested.

They can also ask where human approval occurs. Full automation is not automatically better when a task affects money, personal data, physical devices, or regulated decisions.

Model providers face another pressure. If value moves toward harnesses and outcomes, access to a strong foundation model becomes only one component of the product.

Agent platforms can differentiate through domain knowledge, tool integrations, evaluations, and workflow design. That can reduce the commercial advantage of owning the model layer.

At the same time, model companies can move upward. They can offer managed agents, hosted execution, persistent sessions, and specialized developer systems.

This creates a contest over who owns the relationship with the customer. Infrastructure providers want developers to build on their platforms, while agent companies want buyers to pay for completed business work.

Device manufacturers form a third group. Phones, cars, and smart glasses already control important user interfaces, sensors, identities, and permissions.

Embedding agents into those devices can give manufacturers a distribution advantage. It also makes them responsible for agent behavior close to the user and, sometimes, the physical world.

Beijing’s industrial policy provides a preview. Its industrial agent plan targets 100 high-quality industrial datasets, 50 intelligent solution providers, and 100 advanced industrial agents by 2028.

That plan emphasizes the full lifecycle, including development, deployment, monitoring, behavioral auditing, tuning, and iteration. These are harness concerns even when the document uses different terminology.

Manufacturing also makes outcome measurement more concrete. An industrial agent can be assessed against downtime, defect detection, process stability, energy use, or verified task completion.

However, physical operations raise the cost of mistakes. An incorrect answer in a chat window is different from an incorrect control recommendation on a production line.

The more an agent can act, the more its harness must restrict, record, and verify those actions. This creates demand for permissions, simulation, rollback, and human escalation.

Developers should expect buyers to request evidence for these controls. A polished demonstration will carry less weight than logs, test results, failure boundaries, and recovery procedures.

The Beijing AI agent policy therefore supports agent adoption while raising the engineering bar. It treats operational reliability as an industry layer that companies must design and defend.

OPC Support Tests the One-Person AI Company Model

Beijing is pairing agent infrastructure with a bet that very small companies can produce commercially significant software.

OPC means One Person Company in this policy context. It describes an AI-driven business operated by one founder or a very small team, not a new corporate legal form.

The concept rests on a straightforward claim. Agents can perform enough coding, research, design, marketing, operations, and customer support to reduce the staffing required for a viable company.

Beijing had already moved beyond rhetoric before the new agent measures appeared. Its official OPC action plan, issued in June, supports businesses focused on agent development, generated content, model tuning, and industry applications.

The plan calls for dedicated startup services, community workspaces, model access, computing resources, data support, financing assistance, competitions, compliance guidance, and university programs.

It also allows credit evaluation to consider token consumption, competition results, founder credit, intellectual property, and self-developed technology.

That provision reveals both ambition and risk. Token consumption can indicate activity, but activity is not revenue, product quality, or customer retention.

A company can consume substantial model capacity while searching for a workable product. Another can build a valuable service with fewer tokens because its workflow is efficient.

The reported shift toward value-based billing partly corrects that tension. If Beijing wants outcome-oriented agent businesses, support programs should measure accepted work and durable demand, not only resource use.

OPCs could benefit more than larger companies from mature harnesses. A solo founder cannot manually supervise every agent action across development, support, sales, and administration.

The founder needs repeatable workflows, clear permissions, durable context, and alerts for exceptional cases. Otherwise, automation increases the number of failures requiring personal attention.

This makes OPCs a practical test for agent reliability. A system that works only with constant engineering oversight cannot replace the functions of a larger team.

Beijing has also supported local OPC communities and public showcases. Official accounts describe AI businesses working in education, industrial applications, entertainment, animation, health, and personal data visualization.

These cases show the range of possible products, but they do not yet establish the model’s general success. Showcase selection naturally favors compelling examples.

The economic test will be whether OPCs retain paying customers, manage compliance, handle support demands, and survive changes in model pricing or platform access.

Concentration risk deserves attention. A solo company can depend heavily on one founder, one model provider, and several external APIs.

A policy-backed community can reduce startup friction, but it cannot remove that dependency. Service interruptions, account restrictions, or model changes can affect the entire business.

Liability is another open issue. If an autonomous agent sends incorrect financial information or mishandles protected data, the company remains accountable.

The founder must maintain review systems even when agents perform most operational tasks. Small headcount does not reduce legal obligations.

Security also becomes harder when several agents share broad access. An instruction hidden inside a document can manipulate an agent unless the harness separates untrusted content from executable commands.

OPCs need narrow permissions, credential isolation, action logs, and approval gates. These controls can feel burdensome, yet they protect the founder from automation operating beyond its mandate.

Universities and startup communities can help by teaching operational discipline alongside prompt design. Beijing’s OPC plan already calls for courses combining large-model engineering with entrepreneurship.

The strongest version of the OPC model is not one person asking a chatbot to run a company. It is one accountable founder managing a tested system of specialized agents.

That distinction will determine whether OPC becomes a durable business form or a temporary label attached to conventional freelance work.

The Policy’s Biggest Promise Is Also Its Biggest Risk

Beijing wants agents to act across more devices and industries, but broader agency increases the cost of unreliable behavior.

China’s national framework already defines agents as systems capable of perception, memory, decision-making, interaction, and execution.

A May 2026 agent governance framework places safety, reliability, and trustworthiness across development, deployment, and promotion.

Beijing’s growth policy must operate within that governance direction. Support for rapid adoption does not remove requirements concerning data, algorithms, security, consumer protection, or accountability.

The reported measures create several unresolved questions. First, the authoritative document and any implementation guidance need publication in official databases.

Secondary reports provide a detailed vocabulary list, but developers need exact definitions. A policy term can affect eligibility, reporting obligations, procurement, and access to support.

Second, value-based billing needs verifiable metrics. The government can encourage the model, but customers and providers must negotiate what value means for each task.

Simple tasks allow direct acceptance tests. Complex research, creative work, strategic advice, and long-term automation do not.

Third, interoperability requires security boundaries. Agent-to-agent communication can improve coordination while creating new paths for impersonation, data leakage, and unauthorized action.

A protocol needs authentication, consent, revocation, provenance, and auditability. Message compatibility alone is not enough.

Fourth, device integration requires clear permission models. An agent embedded in glasses or a car can access sensitive contextual information and affect user safety.

Developers should expect stronger scrutiny when software moves from suggesting actions to executing them. Every added tool expands both usefulness and potential harm.

Fifth, OPC programs need outcome data. Startup counts and token consumption can show participation, but they do not prove sustainable businesses.

Useful measures would include customer retention, recurring usage, task acceptance, compliance incidents, founder workload, and survival without continuing support.

The policy also creates a procurement challenge. Government and enterprise buyers can stimulate the market by purchasing agent services, but procurement rules often assume conventional software.

Outcome contracts need methods for evaluating probabilistic systems. Buyers must specify tolerances, escalation procedures, data handling, and responsibility for downstream consequences.

The risk is not that Beijing has named fashionable terms. The risk is that institutional adoption moves faster than the controls needed to make those terms operational.

Yet waiting for perfect reliability is not realistic. Agent systems improve through deployment, observation, and correction.

The practical answer is bounded deployment. Organizations can begin with reversible tasks, limited permissions, strong logging, and human review.

They can expand autonomy only when evidence supports the change. This mirrors how reliable harnesses decompose long tasks and evaluate progress instead of asking one agent to complete everything at once.

The policy’s most valuable contribution may be making these engineering questions visible to nontechnical decision-makers. Once Harness Engineering appears in a policy framework, reliability becomes a procurement and management issue.

That visibility can improve the market if buyers demand evidence. It can also encourage superficial relabeling if companies add “harness” to product descriptions without changing their systems.

The verification gap around the new document therefore matters. Policy language should be evaluated through the official text, subsequent rules, funding criteria, technical standards, and procurement documents.

Readers should watch three signals over the next three months.

The first is official publication of the complete measures and their implementing agencies. Exact text will show whether the reported concepts carry definitions, deadlines, or enforceable programs.

The second is a concrete pilot for value-based agent billing. A credible pilot should disclose acceptance criteria, failure handling, and how providers prove task completion.

The third is technical guidance for agent interoperability and device permissions. Standards, reference implementations, or conformance testing would turn broad support into usable infrastructure.

These signals will test the article’s central judgment. Beijing is trying to shift AI economics from selling computation toward selling accountable work.

If implementation focuses mainly on token credits and startup counts, that judgment weakens. The policy would remain an expansion of familiar industrial support under newer language.

If procurement and technical rules reward verified outcomes, controlled execution, and interoperable agents, the shift becomes substantive.

For developers, the immediate lesson is to treat the harness as product architecture. Record actions, constrain tools, test outcomes, and preserve evidence before expanding autonomy.

For enterprise buyers, the lesson is to separate agent activity from business value. Ask how success is measured and how failure is detected.

For knowledge workers, the policy signals that agents will increasingly enter everyday devices and workflows. Maintaining a searchable personal knowledge system can improve context while keeping source material available for review.

The Beijing AI agent policy has put advanced industry language into a public economic agenda. The next question is whether Beijing can turn that language into measurable, accountable agent work.