9月7日，国内首份《中国办公Agent用户行为不完全报告》在北京发布。这份基于LobsterAI百万级用户数据的报告，为长期处于「黑箱」的办公Agent市场投下了第一束系统性的光。此前我们对办公Agent的认知大多来自厂商宣传，但真实用户到底在用Agent做什么？用得多深？在什么时间用？这些最基本的问题，始终缺乏数据答案。
报告发布时机也颇有深意。两个月前，北京市发改委等四部门联合印发《北京市关于加快智能体引领发展的若干措施》，明确推动代理式人工智能技术创新。政策顶层设计与用户真实行为，正在北京形成有趣的共振。
地理版图：北京领跑，但62.79%的用户不在一线城市
报告最直观的发现是用户地理分布「头重脚不轻」的格局。北京以显著优势位居全国城市用户量榜首，与上海、杭州、广州、深圳五城合计占比约27%。作为全国软件与IT行业最密集的城市，北京成为智能体经济深度实践的核心力量，并不令人意外——最早拥抱新技术的，永远是制造新技术的人。
但真正值得关注的数字是62.79%。这是排名前十城市之外的用户占比。它意味着办公Agent早已突破一线城市壁垒，正在向更广阔的地域渗透。从新一线城市到三四线城市，越来越多的职场人开始将Agent纳入工作工具箱。这是比「北上广深杭领跑」更重要的信号——Agent不是少数极客的玩具，它正在成为普惠的生产力工具。
放眼全球，海外用户占比达到12.75%。美国以2.73%的份额领跑海外市场，澳大利亚、越南、马来西亚、日本等国也形成了颇具规模的用户群体。一款中国本土的办公Agent产品，未经大规模海外推广就自然渗透到全球超一成用户，这至少说明办公Agent的需求是全球性的。
使用深度：20%用户消耗87.4%算力，付费是结果不是原因
如果说地理分布描绘了Agent的「广度」，使用深度数据则揭示了Agent的「浓度」。
报告显示了一个极其显著的「头部效应」：前20%的用户消耗了87.4%的算力，其中前5%的用户独占53.5%的token消耗。这比互联网常见二八定律更为极端——不是20%用户贡献80%价值，而是20%用户几乎「霸占」了整个系统的算力资源。
这种极端分布背后，是办公Agent与传统SaaS产品本质上的不同。传统办公软件的使用强度有天然上限——你一天最多工作十几个小时。但Agent不一样，它可以并行处理任务，可以无人值守运行，可以在你睡觉时继续工作。对于重度用户来说，Agent不是「打开—使用—关闭」的工具，而是7×24小时运转的「数字员工」，使用越深，越能发现更多能力边界，形成正向循环。
付费用户的数据进一步印证了这一点。付费用户在token消耗、任务量和月活跃天数上，分别达到免费用户的6.2倍、5.2倍和3.0倍。关键因果关系需要厘清：不是付费了所以用得多，而是用得多了所以愿意付费。重度使用是付费的驱动力。
时间革命：近60%算力消耗在非办公时间
如果说这份报告只有一个数据值得记住，那应该是这个：近60%的token消耗发生在常规办公时间之外。
工作日晚间占34.3%，周末占25.6%。也就是说，传统意义上的「下班时间」和「休息日」，反而成了办公Agent最忙的时段。24小时使用曲线呈现上午10点与下午15至16点的双高峰——这符合正常的工作节律——但晚22点的小幅回升和凌晨0点仍有3.7%的调用量，才是真正值得玩味的信号。
这背后是两种完全不同的人机协作模式。白天，人和Agent一起工作，人发出指令，Agent即时执行，这是「有人值守」模式；到了晚上和周末，人去休息了，Agent继续干活，这是「无人值守」模式。报告数据为第二种模式提供了有力佐证：12.2%的模型调用并非由人实时发起，其中定时任务的单条指令平均带动64次模型调用，是人工任务（10.6次）的6倍之多。
「真正的生产力革命，不是你工作的时候AI帮你更快，而是你睡觉的时候AI还在替你工作。」
「无人值守」模式的兴起，正在从根本上重新定义「工作时间」。工业时代，工作时间是工人在工厂的时间；信息时代，是白领坐在电脑前的时间；智能体时代，工作时间可能意味着——只要你的Agent在运行，工作就在进行。
任务进化：从对话到做事，单次任务规模5个月增长3.1倍
办公Agent的用户行为正在经历一场静悄悄的「重量升级」。
报告显示，单次任务规模在5个月内增长了3.1倍，仅8月环比增幅就高达53%。这个数字的重要性怎么强调都不为过——用户对Agent的信任度正在以惊人的速度提升。一年前，大多数人还只敢让Agent写写邮件、查查资料；现在，他们开始把更复杂、更长周期、更有风险的深度工作托付给Agent。
从任务类型来看，泛「编程」与调试以38.5%的占比遥遥领先。这个结果初看在意料之中——毕竟Agent的技术底子就是大语言模型，写代码本来就是它的强项。但深入一层想，背后有一个更有意思的现象：很多不会编程的人，现在也开始通过编程的方法来解决自己的问题。做市场的可以让Agent写爬虫抓取竞品数据，做设计的可以让Agent写脚本批量处理图片，做运营的让Agent写自动化程序处理报表。Agent降低了编程的门槛，让「编程思维」从程序员的专属技能，变成了普通人也能掌握的问题解决方法论。
模型生态：Agent成国产新模型最快落地通道
在模型调用层面，DeepSeek V4 Flash以52.2%的占比位居榜首，如果加上其视觉实验版本，合计占比超过75%。但真正的看点在于新模型的渗透速度。
报告指出，Agent正在成为国产新模型触达真实工作负载最快的渠道——新模型上线3天内，用户渗透率就接近50%。传统模式下，新模型发布后，开发者需要花几周甚至几个月去适配、测试、集成。但在Agent平台上，模型即插即用，用户只需切换一下选项，就能立刻体验最新的模型能力。
这也意味着，Agent正在成为国产大模型竞争的新战场。谁能在Agent场景下表现更好——更快的响应速度、更低的调用成本、更强的任务执行能力——谁就能抢占这个增长最快的模型分发渠道。
作为国内大厂首个开源桌面级Agent，LobsterAI的崛起是这个时代的一个缩影。代码100%开源、能力透明可查、可连接文件/终端/浏览器/本地项目、支持多IM远程指挥、13套岗位专家技能包。GitHub超6k Stars、8月用户数突破100万、跻身国内办公Agent「四强」，这些成绩证明了开源路线在Agent领域的可行性与爆发力。
中国办公Agent已经进入「重任务、深集成、广覆盖」的新阶段。这不是渐进式演变，而是一场结构性跃迁。当你明天早上醒来，你的Agent可能已经替你做完了今天一半的工作——这听起来像科幻，但数据告诉我们，它正在发生。
明天见。
On September 7, China's first Incomplete Report on Office Agent User Behavior was released in Beijing. Based on LobsterAI's data from one million real users, the report casts the first systematic light on the office agent market, which has long operated as something of a black box. Until now, our understanding of office agents has mostly come from vendor announcements. But what are real users actually doing with agents? How deeply do they use them? When do they use them? These most basic questions have consistently lacked data-driven answers.
The timing is also telling. Two months prior, four Beijing municipal agencies jointly issued Several Measures on Accelerating Agent-Led Development, explicitly pushing for agentic AI innovation. Top-level policy design and real user behavior are forming an interesting resonance in Beijing.
Geographic Landscape: Beijing Leads, but 62.79% of Users Are Outside Tier-1 Cities
The most straightforward finding is the geographic distribution pattern — top-heavy but not exclusively so. Beijing ranks first nationwide with a significant margin, and together with Shanghai, Hangzhou, Guangzhou, and Shenzhen, these five cities account for approximately 27% of users. As China's densest hub for software development and IT, Beijing becoming the core force of the agent economy is no surprise — those who earliest embrace new technology are always those who build it.
But the truly noteworthy figure is 62.79%. That's the share of users outside the top 10 cities. It means office agents have long broken through the tier-1 city bubble and are penetrating much wider geographies. From new tier-1 cities to third- and fourth-tier cities, more professionals are adding agents to their work toolboxes. This is a more important signal than "the big five leading the pack" — agents aren't toys for a small number of geeks. They're becoming a democratized productivity tool.
Looking globally, overseas users make up 12.75% of the total. The United States leads overseas markets at 2.73%, while Australia, Vietnam, Malaysia, Japan, and others have also formed sizable user groups. A China-grown office agent product, without large-scale overseas promotion, has naturally reached over 10% of global users — suggesting at minimum that demand for office agents is global.
Usage Depth: 20% of Users Consume 87.4% of Compute
If geographic distribution describes the "breadth" of agent adoption, usage depth reveals its "intensity."
The report shows an extremely pronounced "head effect": the top 20% of users consume 87.4% of all compute, with the top 5% alone accounting for 53.5% of token consumption. This is even more extreme than the familiar 80/20 rule — it's not that 20% of users contribute 80% of value, but that 20% virtually monopolize the system's compute resources.
Behind this extreme distribution lies a fundamental difference between office agents and traditional SaaS. Traditional office software has a natural ceiling on usage intensity — you can work at most a dozen hours a day. But agents are different. They process tasks in parallel, run unattended, and keep working while you sleep. For power users, an agent isn't a tool you "open — use — close"; it's a 24/7 "digital employee." The deeper users engage, the more they discover what agents can do, creating a virtuous cycle.
Paying user data further confirms this. Paid users reach 6.2x, 5.2x, and 3.0x the levels of free users in token consumption, task volume, and monthly active days. There's a critical causal relationship to clarify: people don't use more because they pay — they pay because they use more. Heavy usage drives payment, not the reverse.
The Time Revolution: Nearly 60% of Compute Happens Off-Hours
If there's one statistic from this report worth remembering, it's this: nearly 60% of token consumption happens outside regular business hours.
Weekday evenings account for 34.3%, and weekends for 25.6%. In other words, traditional "off-hours" and "rest days" have become the busiest periods for office agents. The 24-hour usage curve shows twin peaks at 10 a.m. and 3–4 p.m. — consistent with normal work rhythms — but the slight rebound at 10 p.m. and the persistent 3.7% call volume at midnight are the signals truly worth savoring.
Behind this lie two entirely different modes of human-AI collaboration. During the day, humans and agents work together — humans issue commands, agents execute in real time. This is the "attended" mode. At night and on weekends, humans rest and agents keep working. This is the "unattended" mode. The report provides strong evidence for the second mode: 12.2% of model calls are not initiated by a human in real time, and scheduled tasks average 64 model calls per instruction — six times the 10.6 calls of manual tasks.
"The real productivity revolution isn't AI making you faster while you work — it's AI working for you while you sleep."
The rise of "unattended" mode is fundamentally redefining "work hours." In the industrial age, work hours meant time in the factory. In the information age, work hours meant time at the computer. In the agent age, work hours might mean — as long as your agent is running, work is happening.
Task Evolution: From Conversation to Execution, 3.1x Growth in 5 Months
Office agent user behavior is undergoing a quiet "weight upgrade."
The report shows single-task scale has grown 3.1x in five months, with a 53% month-over-month increase in August alone. The importance of this number cannot be overstated — it means user trust in agents is improving at a remarkable pace. A year ago, most people only dared let agents write emails or look up information. Now, they're entrusting agents with more complex, longer-cycle, higher-stakes deep work.
By task type, "programming" and debugging broadly defined leads by a wide margin at 38.5%. At first glance, this is expected — agents are built on LLMs, and coding has always been their strong suit. But looking deeper, there's a more interesting phenomenon: many people who can't code are now solving problems through programming methods. A marketer can ask an agent to write a web scraper for competitor data. A designer can ask for batch image processing scripts. An operations professional can ask for report automation. Agents lower the barrier to programming, turning "computational thinking" from an exclusive programmer skill into a problem-solving methodology accessible to everyone.
Model Ecosystem: Agents Become the Fastest Channel for New Domestic Models
On the model front, DeepSeek V4 Flash ranks first with a 52.2% share, and including its vision experimental version, the combined share exceeds 75%. But the real story is penetration speed.
The report notes that agents are becoming the fastest channel for new domestic models to reach real workloads — within three days of launch, user penetration approaches 50%. Traditionally, after a new model launches, developers spend weeks or months adapting, testing, and integrating before users benefit. But on agent platforms, models are plug-and-play — users simply switch an option and immediately experience the latest capabilities.
This also means agents are becoming a new battleground for domestic large model competition. Whoever performs better in agent scenarios — faster response, lower cost, stronger execution — will capture this fastest-growing distribution channel.
As the first open-source desktop-level agent from a major Chinese tech company, LobsterAI's rise is a microcosm of this era. 100% open-source code, transparent capabilities, connectivity to files, terminals, browsers, and local projects, support for remote command via multiple IM apps, and 13 role-based expert skill packages. With over 6k GitHub stars, 1 million users as of August, and a place among China's office agent "Final Four," these achievements validate the open-source route's feasibility and explosive potential.
China's office agents have entered a new phase of "heavy tasks, deep integration, broad coverage." This isn't incremental evolution — it's a structural leap. When you wake up tomorrow, your agent may have already done half of today's work for you. It sounds like science fiction, but the data tells us it's happening.
See you tomorrow.
Geographic Penetration · Head Effect · Time Revolution · Unattended Mode · Task Weight Growth · Open-Source Agent
