Beijing's 10-Point Plan for an Agent Economy
Beijing has unveiled a sweeping policy document titled "Several Measures for Accelerating the Development of Intelligent Agents in Beijing," laying out a ten-point blueprint for what it calls the "intelligent agent economy." The document, released today, is packed with cutting-edge terms like Agentic AI, Harness Engineering, AI OS, FDE, OPC, Token Economy, and more—terms that until recently were mostly confined to academic papers and tech jargon. Now, they're official government policy.
Point one focuses on foundational models, with a clear directive: "continuously improve the actual task completion ability of large models." The reasoning is simple but brutal. If an agent performs 100 steps with 99% accuracy per step, the chance of getting everything right is only about 36.6%. Even at 99.9% accuracy, it's just 90.5%. So the real value of an agent is a multiplication: model capability × long-term reliability × environmental operability × result verifiability. If any factor approaches zero, the whole thing collapses. That's why code became the first area where agents exploded—it's digital, rollback-able, and verifiable.
Point two introduces "Harness Engineering"—the art of building a robust system around the model. It's about context engineering, task persistence, multi-agent collaboration, and system scalability. A paper from April showed that by improving tools, middleware, and long-term memory (without touching the base model), Pass@1 scores on Terminal-Bench2 jumped from 69.7% to 77.0%. The policy also mentions skill markets and software stores, hinting that future competition won't just be about app rankings but about how agents discover and adopt skills.
Point three pushes for restructuring software architecture around agents, with the core concept being "demand intelligence." This means moving beyond simple functional or process intelligence to systems that understand what you want and figure out the steps themselves. It also calls for FDEs (Frontier Deployment Engineers) who go to customer sites, break down processes, connect systems, and clean data. The policy identifies six benchmark scenarios: science, healthcare, education, government affairs, manufacturing, and culture.
Point four moves from software to hardware, embedding agents into phones, glasses, earphones, wearables, robots, and cars. The goal is "five-in-one integration of chip, model, cloud, and terminal usage." Edge devices handle low-latency tasks, while the cloud manages complex planning. Task state sharing across devices means you can start a task on earphones, continue on a phone, and finish on a car.
Point five is the most exciting for individuals: supporting "One-Person Companies" (OPC). The idea is that agents lower execution and coordination costs, allowing one person to do the work of a team. The policy backs this up with flexible computing, incubation, guidance, finance, and community support. For individuals, the key assets are industry judgment, customer trust, and a reusable agent system.
Point six dives into the Token economy. Agents can burn through 2 billion tokens without achieving anything, so tokens are better as a cost measure than a value currency. The policy encourages moving from billing by token consumption to value-based billing. Three models emerge: TaaS (selling tokens), AaaS (selling agent capabilities), and RaaS (selling results). The higher up, the closer to customer value, but also the harder. The document even proposes token coupons and service coupons to stimulate demand.
Point seven addresses security. Agents have action rights—they can edit code, initiate payments, control devices, and communicate for users. Risks expand from "saying something wrong" to "doing something wrong." The policy moves from content review to runtime governance, with categorized regulation: low-risk tasks can be tested quickly, while high-risk tasks require strict inspection.
Point eight launches the "Galaxy Computing Corridor" project, building a new computing infrastructure for high-frequency, low-latency agent demands. It integrates 5G-A, 6G, and F5G, and taps into existing computing power. The document also supports financial products for agent companies, whose most valuable asset is their "task trajectory"—the complete chain of goals, plans, actions, and results.
Point nine promotes open source and openness, including a China-South Asian Cooperation Center for AI applications and a globally influential open-source community. The agent era thrives on interconnectedness, so protocols and open source are key.
Point ten provides guarantees, coordinating national and municipal funds to support technical breakthroughs and demonstration applications.
Key Points
- Ten measures cover everything from foundational models to security and open source.
- Harness Engineering is a central concept: building robust systems around models.
- One-Person Companies are officially supported, lowering the barrier to entrepreneurship.
- Token economy shifts from billing by tokens to billing by value.
- Security moves from content review to runtime governance with categorized regulation.
- Galaxy Computing Corridor aims to build a new computing infrastructure for agents.
