Tencent tells of 90 billion tasks done by its WorkBuddy agent

At the 2026 Tencent Global Digital Ecosystem Summit, Tencent told of 90 billion tasks done by its WorkBuddy tool since March. Its enterprise version is in more than 50 industries and is at 50% of the Fortune 500 companies in China. Leaders of 3 companies tell that the next problem is the cost of tokens, small units of text that a model reads.

Where cost comes from

Chen Liang of Tencent Cloud tells that software has a lower cost for each unit when it has more users. Model use can cost almost the same for each new user, because more users cause more tokens. Companies try open-source models from China and short input text.

The numbers

Company data:

• Xiaoyuzhou, a podcast app: more than 70% of frequent tasks use the cache.
• Xiaoyuzhou input cost is as low as 0.02 yuan ($0.003) for each task.
• Youka Games: cost for each inference decreased about 30%.
• Taomee Network: content output increased by almost 80%.

What is not known

Ding Junwei of Tec-Do tells that agents can do more than 90% of advertising tasks. Li Tianyu of YuanCe Future tells that large models do not stay stable at the 99.99% level that robots must have. No external test shows these data.

Insight: The limit is now human approval, not output

Tec-Do tells that a company can use only about 100 of 1,000 assets, because a person selects the budget. Mintlify tells that only 9% of the companies in its report let agents release changes without a human check. Agents make output at low cost, and human approval and budget become the limit. The next move to look for: Tec-Do will give campaign results to its model, possibly to select good assets.