Operating decision relevance
Beijing's new token economy plan names the software needed to put AI agents to work: agent memory, permissions, communication between components, model routing and developer tools.
For a small AI company, that is a useful place to start reading.
The municipal action plan, published on September 18 and running through 2028, describes the infrastructure and software Beijing wants around AI model use. Its scope reaches from inference facilities to the products that turn model calls into work for customers.
For overseas teams building that software, the plan provides a basis for approaching Beijing partners with a specific product and an implementation proposal.
Agent infrastructure is explicitly in the plan
One measure is devoted to the core software behind agents. It identifies memory management, permission controls and communication between components, and calls for the development of agent operating systems and new software-tool companies.
These are problems a team encounters when it tries to make an agent useful beyond a demonstration. The system needs to retain context, access the right tools and data, and carry out a sequence of actions reliably.
A company with a working product in one of these areas has something concrete to show a potential partner. A useful first demonstration would use the partner's models and a representative business task, with clear results for task completion, execution time and operating cost.
For a specialized team, that makes a focused integration project a sensible starting point for a Beijing operation.
Distribution and evaluation are business opportunities too
The plan also supports platforms that combine model access through a common interface and handle usage measurement and settlement. It connects this with model evaluation and routing: helping customers choose and use the right model for a particular job.
For a developer-tooling company, there are several possible products here. An API gateway could simplify integration across providers. An evaluation service could test models against a customer's actual workflow. Routing software could select a model according to quality, speed and cost requirements.
Those are commercial interpretations of the policy's priorities. They give an overseas team a way to explain what it would contribute to a local platform.
The economic case should use actual provider pricing and measured performance. In a September 28 interview, Beijing economic officials described token prices as being formed through market competition, with government work focused on evaluation, measurement and distribution infrastructure.
There are already organizations to approach
A September 21 policy briefing brought together companies working on different parts of this system.
According to the municipal bureau's account, the China Software Testing Center is working with companies on token-factory grading and performance standards. Qingcheng Jizhi (清程极智) discussed distribution services. Baidu AI Cloud presented platform operations and industry applications, while JD Technology discussed uses in the real economy.
The bureau also encouraged companies to work through Beijing's general AI industry partnership program on joint projects and standards.
For an overseas supplier, these are starting points for a partner shortlist. The relevant conversation depends on the product: evaluation with a testing organization, routing with a distribution platform, or agent deployment with an industry service provider.
Bring a working integration, a concise description of the customer problem and the resources needed to run a pilot. That makes it easier for a prospective partner to decide where the product fits.
International delivery is part of the same program
The plan includes overseas services built around open-model adaptation, agent deployment, developer toolchains and computing. It names multilingual adaptation and overseas operations among the capabilities to develop.
This is particularly relevant to a foreign team that understands a market Beijing companies want to serve. Technical integration and local customer knowledge can sit in the same partnership: adapting an agent to a language, connecting it to business systems and supporting its deployment with customers.
The immediate task is to define that contribution. Specify the workflow, the models and infrastructure it uses, the integration work required and who will operate the service. With those pieces in place, a Beijing software partner has a project it can evaluate and build around.