When agents start spending: WeChat Pay's AI card and the rollout of Chinese agents

Agents in China are moving from chatting to executing tasks. On August 31, WeChat Pay announced an update to its AI card, adding support for DeepSeek Harness and OpenClaw after WorkBuddy and QClaw. Once authorized, users state their needs in the chat. The agent then recommends items, places the order, and processes the payment.

The AI card currently supports paid calls to over 700 Pay Skills on Skillhub via DeepSeek Harness and OpenClaw. More scenarios are planned. Activation is conversational: users copy the install command to the agent, which handles download, installation, and configuration in one step. During the first payment, the agent generates a binding link.

The authorization design of the payment chain is critical. Each order triggers a payment authorization prompt that requires secondary confirmation on the phone before the AI card is charged. Only then can the agent invoke the Pay Skill to complete the task. Tencent compares the card to the money given to someone running an errand: the user sets and manages the limit, and the card operates in a strong authorization mode.

Regarding security, WeChat restated three boundaries. The AI card is fully isolated from the main WeChat Pay account, and all spending within an agent can only use the AI card balance. Users set the limit and can adjust it at any time by transferring funds in or out. Without the user's final authorization, the AI cannot spend a single cent.

The significance of this product lies in transforming agents from advisers into executors. For enterprises, an agent that can pay for skills is one that truly integrates into business processes. DeepSeek also designates WorkBuddy (including CodeBuddy) and OpenCode as official partners, stating they have fully integrated V4.1 Flash.

The limits of the rollout still require observation. Agents autonomously calling tools and payments raise higher demands for permission management, auditing, and responsibility allocation. WeChat uses isolated accounts and per-order confirmation to contain risk. This engineering approach may become a reference for agent commercialization in China.

This model assumes agents can reliably call external tools and payment interfaces. This is precisely the scenario where models like V4.1 Flash continue to optimize for long context and cache costs.