Beijing's latest policy directive is more ambitious than anticipated. Recently, four municipal departments, including the Beijing Municipal Development and Reform Commission, jointly released a significant policy document titled "Several Measures to Accelerate the Development of Agentic Intelligence in Beijing." The document is concise, consisting of just ten articles. Why does it surpass expectations? Because terms like Agentic AI, Harness Engineering, AIP, AI OS, FDE, OPC, and Token Economy—previously confined to product proposals, technical discussions, and cutting-edge research papers—have now entered a formal policy document. This is exceptionally rare.

This development has sparked widespread discussion. Some view it as a blueprint for the future agentic economy, while others believe it will directly impact every industry and individual. So, what exactly does Beijing's new agentic policy entail? And what implications does it hold for us? Let's break it down article by article.

Article 1: Continuously Enhancing Foundational Model Capabilities

What does the first article mean? Simply put, it's about making models not just smarter but also more effective at performing tasks. The policy states: "Continuously improve the ability of large models to complete real-world tasks, supporting innovative entities in tackling common technical challenges such as tool invocation, long contexts, multi-agent collaboration, and memory." In the past, model quality was judged by benchmarks and rankings—how many Olympiad problems were solved, what the code pass rate was, and which competitor was surpassed. However, being good at answering questions is different from being good at doing work. For example, early companies often asked puzzle questions in interviews, like how long a rope burns or when a frog climbs out of a well. Later, they stopped because the answers didn't correlate with a candidate's ability to manage a three-month project independently. What companies really want to know is whether a candidate will drop the ball, call for help when issues arise, and deliver complete results on time. Now, we have new standards: Can an agent complete a task from start to finish in a real environment with permissions, tools, anomalies, and deadlines? This is essentially what we expect from employees. How do we ensure an agent meets these expectations? Through tool invocation, long contexts, multi-agent collaboration, memory, and other capabilities written into the policy, such as online learning, continuous learning, autonomous evolution, ultra-long tasks, and complex reasoning. Tool invocation is like using company systems. Long context is like reading a stack of documents in one go. Multi-agent collaboration is like working with others. Memory is like remembering tasks from last week. These technical terms are simply what a person needs to do their job. So, next time you choose an AI tool, don't just look at its ranking. Pick a daily task in your company, have it perform it 20 times, and see how often your business manager signs off. Rankings test intelligence, but business tests work capability. Your money ultimately pays for results, not smarts.

Article 2: Strengthening Core Foundational Agentic Technology

The second article is lengthy and technical, but it addresses a straightforward concept: infrastructure support. The policy states: "Support innovative entities in implementing Harness Engineering, focusing on context engineering optimization, task persistence, multi-agent coordination, and system scalability to solidify the agentic common foundation." Harness Engineering refers to the peripheral system around the thinking model—how to organize information, ensure it gets the right amount of data, remember previous steps, resume from interruptions, coordinate multiple agents without conflicts, and catch errors to prevent cascading failures. The harness determines whether the model can complete tasks effectively. As validated in April by "Agentic Harness Engineering," without changing the model, 10 rounds of automatic evolution on the harness improved the first-pass rate on a terminal task benchmark from 69.7% to 77.0%. Why? Because the model is the engine, and the harness is the transmission, chassis, brakes, and steering wheel. While the world has focused on engine horsepower, it's now time to build the car. Thus, the policy also promotes supporting elements like skill markets, agent interconnection protocols, development platforms, and software stores. For you, this means your true AI assets might not be the specific model used, but the records of failures—what tasks AI botched, at what step, what inputs caused crashes, what outputs passed inspection, and how past accidents were handled. These are the moats others can't steal.

Article 3: Accelerating Native Agentic Applications and Benchmark Scenario Construction

The third article, covering science, healthcare, education, government, manufacturing, and culture, has one core objective: integrating agents into real work, not just demonstrations. How? First, AI-native software: Encourage innovation entities to rebuild underlying architectures and interaction paradigms based on agents, developing native AI software and super software driven by intelligent demand. Adding a big screen and voice assistant to a fuel car doesn't make it an electric vehicle, which is redesigned from the chassis up. So, rebuild the foundation, not just add features. Second, AI OS: Promote the development of enterprise-level and industry-level artificial intelligence operating systems to break ecological barriers, enhancing cross-device coordination, cross-ecosystem development, and cross-scenario deployment. "Operating system" implies a base for all agents to run on, like Windows or Android. Third, FDE (Frontier Deployment Engineers): Place engineers in customer business environments, observe their workflows, help them complete initial tasks, and feed insights back into product development. Additionally, the policy mentions "building a precise match platform for scenario supply and demand, organizing the release of real business scenario requirement lists for agents." This means bringing supply and demand together with a matchmaker. If you're in science, healthcare, education, government, manufacturing, or culture, this article deserves your attention. However, policy support doesn't equate to industry access; procurement rules and data regulations still apply.

Article 4: Promoting Integrated Development of Smart Terminals and Agents

The fourth article translates to bringing AI out of the screen. The policy states: "Enhance agents' environmental perception, autonomous decision-making, task execution, and terminal deployment capabilities, supporting innovation entities in deeply embedding agent capabilities into smart terminals, developing next-generation smartphones, smart glasses, smart earphones, wearable devices, smart robots, and smart cars." These devices have sensors, are on-site, and can move. Today's agents are like capable advisors locked in a room—they answer questions eloquently but can't go downstairs to handle tasks because they lack eyes, hands, and legs. So, open the door, give them eyes, and attach wheels. Specific steps include promoting coordinated development of chips, models, clouds, terminals, and applications; supporting joint hardware-software definition between terminal makers and model companies; developing core hardware components for embodied intelligence; and creating benchmark "Beijing-Made" agent terminals. Previously, model companies, phone makers, and chip firms worked independently, resulting in functional but awkward integrations. Now, they must collaborate from the start, deciding on chip computing power, model size, local vs. cloud processing, and more. This shift in industrial division will have a greater impact than device sales methods.

Article 5: Supporting New Innovation Models like OPC (One Person Company)

The fifth article introduces the popular concept of OPC, or One Person Company. Why is OPC trending? Economist Coase argued that companies exist because buying services on the market is costly—negotiation, contracts, oversight, and risk management create transaction costs. When these costs are high, it's better to hire inside an organization. The company's boundary is defined by transaction costs. As market purchases become less cumbersome, companies shrink. OPC is the logical outcome. It's not a new legal entity—one-person businesses have long been legal. Responsibilities like shareholder liability, corporate independence, taxes, accounting, intellectual property, and data compliance remain. However, operational support can be provided. Thus, the policy includes OPC communities, full-cycle service stations, elastic computing supply, professional incubation, entrepreneurship coaching, and tech finance. The vocabulary itself signals a shift.

Article 6: Encouraging the Development of Token Economy

The sixth article addresses a fundamental business question: How to charge? The policy introduces three concepts: TaaS (Token as a Service), selling resources like electricity; AaaS (Agent as a Service), selling capabilities like a job's output; and RaaS (Results as a Service), selling outcomes and associated risks like a contractor. The policy encourages moving from token consumption billing to value billing to improve intelligent economic efficiency. Token consumption billing is like paying a nanny based on electricity usage—she runs appliances all day, but the meal is poorly cooked and the house is dirty. Value billing focuses on whether the meal is done, regardless of electricity used. However, implementing value billing is challenging because defining "task completion" is difficult. For example, an AI customer service company charges per resolved issue but has a rule: if the user returns to the same conversation, the payment is voided. Without this, user silence would equate to resolution, even if the user just gave up. Currently, the policy encourages, not mandates, value billing, but the ultimate aim is results.

Article 7: Comprehensively Enhancing Safety Governance Capabilities

The seventh article addresses risks that emerge when agents actually work. Previously, AI risks involved hallucination or fabricating references—like an intern writing a flawed report. Now, the intern might have your company's financial system password. The policy proposes mechanisms like: establishing a tiered and classified agent supervision mechanism, building range environments and trusted sandboxes, developing security models to enhance "model-governing-model" capabilities, and offering common services like vulnerability scanning, security testing, attack-defense drills, and security certifications. A practical approach to tiered access is to avoid giving a new intern too many permissions. Use four levels: 1) Read-only assistance (view, suggest, no system access); 2) Draft pending approval (you draft, you approve); 3) Restricted rollback execution (execute within whitelist tools, caps, and time windows, with rollback capability); 4) Limited autonomy (only for high-frequency, low-risk, observable, and reversible tasks with regular audits and budget caps). Also, create a checklist of tasks that always require human approval, such as payments, refunds, contracts, commitments, HR decisions, credit transactions, diagnoses, treatment, production controls, cross-border sensitive data, and irreversible deletions.

Article 8: Enhancing Key Factor Guarantees

The eighth article covers infrastructure, vouchers, and talent. First, roads and electricity: "Establish a multi-level computing power supply system, implement the 'Galactic Computing Corridor' project, accelerate the deployment of new computing infrastructure for high-frequency concurrent, low-latency agent demands, and deeply integrate computing networks with 5G-A/6G and F5G communication networks." Training requires large, long-duration, latency-tolerant computing; agents need fragmented, fast, on-demand computing because a single task may call a model dozens of times, each requiring instant response. The "Galactic Computing Corridor" aims to dispatch computing across regions like a power grid, even aggregating scattered small computing units. Second, vouchers: Computing power vouchers, token vouchers, and agent service vouchers provide low-cost, standardized computing for SMEs and OPCs, lowering usage barriers. Third, talent: "Deepen the professional title system reform, establish a dynamic evaluation system aligned with technology iteration, industry development, and talent growth, and identify urgently needed talents in emerging fields like agent development, large model applications, and data governance." Professional titles signal formal recognition of a profession, with standards, evaluation paths, and career ladders. Roads, electricity, vouchers, and people are the key factors.

Article 9: Promoting Open-Source Development

The ninth article covers two aspects: going global and opening up. First, going global: "High-level construction of the China-SCO AI Application Cooperation Center, study the 'one-country-one-policy' path for large model and agent going global, and strengthen the AI overseas service station to provide policy consultation, compliance training, intellectual property, and risk warning services." This means mapping out common pitfalls in overseas expansion before they occur. Second, opening up: "Establish and improve an open-source contribution evaluation system and support measures, support innovation entities in open-sourcing core codes like autonomous interconnection protocols, agent development frameworks, smart terminal OS, and hardware public boards, and encourage developers to co-build and share open-source ecosystems. Coordinate government investment funds and social capital to support open-source projects." This means contributions to open-source communities can now be quantified, recognized, and financially supported. Why open source? Because in standard wars, open source is the cheapest weapon. The more people use your framework, the more your interfaces, names, and conventions become de facto standards. Eventually, others adapt to your rules, and you become the table itself, not just a competitor.

Article 10: Safeguard Measures

The tenth article addresses leadership, funding, and implementation. Leadership: "The Municipal Development and Reform Commission strengthens overall coordination, working with the Municipal Cyberspace Administration, the Municipal Science and Technology Commission/Zhongguancun Management Committee, and the Municipal Economic and Information Technology Bureau to deepen district-city linkage and government-enterprise collaboration." Funding: "Coordinate the use of national and municipal financial funds, government investment funds, and market-based funds." Implementation: "Organize and implement a batch of key projects, providing support for eligible projects in technology research, common platforms, and demonstration applications." These appear routine, but reports from policy briefings suggest a potential 100 million yuan support for selected projects. However, the original wording is "selected projects receiving maximum support." Why highlight this? Because with major policies, some companies pre-hire and move locations, treating subsidies as guaranteed income, only to find eligibility issues later. The correct approach is to treat policy like an option, not a receivable: prepare but avoid irreversible costs. Then, monitor the intensity of verbs—whether it's "encourage" or "support," "suggest" or "require." Policy builds the road, but you must drive the car.

Final Thoughts

We've now reviewed "Several Measures to Accelerate the Development of Agentic Intelligence in Beijing." Amidst all this content, it boils down to one word: Work. Can agents perform tasks? Can they engage in real work? Who directs their work? How do they get paid for it? What happens when they fail? Where do resources come from? Where does talent come from? This is why the policy exceeds expectations—not in its technical vision, but in its policy scope. Of course, such transformative changes may cause unease. Corporate boundaries are shifting, pricing models are evolving, risk nature is changing, and even professional titles are coming. But don't see this as a wave. It's a road. Though it contains many unfamiliar elements—like OPC's potential or how to measure outcome-based fees—it has a starting point, direction, and signposts. Take it step by step, and you'll understand. Once understood, it's just a task to execute.