NetEase Youdao has released a behavior report based on users of its LobsterAI desktop agent, offering a detailed but provider-specific picture of how office agents are being used in China. The report was published in Beijing on September 7 and summarized through an authorized QbitAI release. It says LobsterAI had passed one million users in August and that its data captures changes in where users are located, what tasks they assign, and when they let an agent work without direct supervision.
The report is not a national census of Chinese office AI. It reflects one product’s users and is published by the product’s operator, NetEase Youdao. That limitation matters. The numbers are useful evidence about a defined service, but they should not be converted into claims about every office agent or every Chinese knowledge worker.
Even with that caveat, the data gives a more concrete view of AI adoption than a product-launch announcement. It suggests that a relatively small group of heavy users consumes most of the computing resources, that tasks are moving beyond simple chat, and that agent activity increasingly happens outside ordinary office hours. The pattern resembles a transition from occasional experimentation to more routine delegation of work.
The report arrives after an earlier wave of workplace-agent products. EastFrontier documented how ByteDance unified office agents under Doubao Work, while Alibaba and Tencent have each tried to build enterprise and productivity use cases. LobsterAI’s contribution is different: it provides product-level evidence of what users actually do after the tools become available.
Usage Is Concentrated in Cities and Heavy Users
LobsterAI’s report says Beijing has the largest concentration of its users. Beijing, Shanghai, Hangzhou, Guangzhou, and Shenzhen together account for about 27% of the product’s user base. At the same time, users outside the top 10 cities make up 62.79% of the total, indicating that the service is not confined solely to the most prominent technology hubs.
The product also has an international component. NetEase Youdao says overseas users account for 12.75% of LobsterAI’s base, with the United States representing 2.73%. Those figures should be read as a snapshot of LobsterAI, not as evidence of China-wide export market share. They do show that a Chinese desktop agent can reach users outside the domestic market without necessarily becoming a global mass-market service.
Resource consumption is even more concentrated. The report says the top 20% of users consume 87.4% of computing power, while the top 5% account for 53.5% of token usage. That distribution has implications for both product design and economics. A provider must serve occasional users cheaply while maintaining enough capacity for the small group that uses the system intensively.
Paid users are especially important in that model. According to the report, paid users use 6.2 times more tokens, complete 5.2 times more tasks, and are active three times as many days per month as free users. These correlations do not establish that paying causes higher use. It may be that people who already have demanding tasks are more willing to pay. The data nevertheless suggests that the strongest revenue potential lies with users who have integrated the agent into recurring workflows.
Longer Tasks Move Beyond a Chat Window
The report says the size of an individual task grew 3.1 times in five months, with a 53% month-on-month increase in August. That is one of its most consequential findings because it suggests users are asking an agent to handle more than short questions or isolated document edits.
LobsterAI reports that nearly 60% of token consumption occurs outside normal office hours. Weekday evenings account for 34.3%, weekends for 25.6%, and midnight still records 3.7% of calls. This may indicate that users are launching tasks before leaving work, reviewing material at home, or setting up automated processes that continue when they are not actively at a desk.
The product’s “unattended” behavior provides another sign of changing use. The report says 12.2% of model calls are not initiated in real time by a person. Scheduled tasks, according to LobsterAI, average 64 model calls per instruction compared with 10.6 for manually initiated tasks. That difference makes sense if a scheduled instruction triggers a longer sequence of searches, file operations, or steps across connected tools.
EastFrontier previously examined concerns among Chinese tech workers about office agents. LobsterAI’s data does not prove that agents are replacing jobs. It does show that some users are delegating increasingly substantial work and allowing systems to operate with less direct supervision. The labor implication is therefore about workflow redesign first, with employment effects still uncertain.
Model Choice and Task Mix Reveal an Agent Market
The report says DeepSeek V4 Flash accounted for 52.2% of LobsterAI model calls, and that the visual experimental version pushed the combined share above 75%. These are not independent measures of DeepSeek’s national market share. They reflect model selection inside one agent product, where price, latency, product integration, and user preferences may all affect the outcome.
The task breakdown is also revealing. General programming and debugging made up 38.5% of tasks, followed by document writing at 10.1% and data and spreadsheet work at 8.8%. Software development and IT accounted for 38.9% of agent work by industry, with design, film, creative work, and internet product operations also represented.
That pattern supports the idea that office agents are not merely chat interfaces. The most common tasks are those where a system can inspect code, generate drafts, manipulate structured data, or coordinate actions across files and tools. LobsterAI says its agent can connect to local files, terminals, browsers, and projects, as well as receive remote instructions through messaging services.
The company’s open-source status is another part of its appeal. LobsterAI’s GitHub page records more than 6,000 stars, and the product says its code is fully open source. Open code does not automatically make an agent safer or better, but it allows developers to inspect a project and adapt it to their own environments.
The report’s lasting value will depend on whether NetEase Youdao releases comparable data over time. One survey can show a moment in adoption. A series of reports could show whether heavy use spreads, whether token concentration declines, and whether agents become more common in nontechnical work. For now, LobsterAI offers a useful, if narrow, empirical window into an agent market that is often discussed through product claims rather than observed behavior.