From drafting emails to finishing the job: ByteDance's Doubao Work ships as a Lark-bound office agent

An office agent is not a smarter chatbot. It is a different kind of tool, one that asks for permission to read the company's chat, open its apps, and finish the work. ByteDance's Doubao Work, released in China on 25 August 2026, is the latest product to make that design choice explicit. The agent is built to inherit the user's existing Feishu workspace as its context. Without that inheritance, the company says, the product does not really work.

Feishu is ByteDance's workplace chat and documents platform, the Chinese counterpart to Slack and Google Workspace combined; outside China the same product is sold as Lark. For a reader who has never opened it, the relevant fact is not the app list but the access it gives: chat history, shared documents, meeting notes, and calendars, all under one permission system. Doubao Work runs on top of that surface. After a user signs in, the agent can read the threads and files it has been granted and use them as the working memory for whatever task it has been handed.

The category matters more than the brand. A "Doubao" in this context is not the consumer chatbot most non-Chinese readers have seen covered; it is the product family ByteDance is using for productivity work, and "Doubao Work" is the new sub-brand for the office agent. The official product page lists content generation and editing across documents, slides, tables, reports, research summaries, and data analysis, plus image, video, web, and app generation. The desktop client is downloadable from the official site, and ByteDance is offering a 30-day free subscription tied to download, sign-in, or upgrade.

The interesting bet is not the feature list. It is the coupling. Two design choices set this product apart from a chat model wrapped in browser automation. The first is what the company calls frame-selection editing: instead of regenerating a whole paragraph when the user wants one clause changed, the agent can rewrite only the selected span, a small thing that the company says saves both compute and the user's patience. The second is permissioned autonomy. With explicit user authorization, Doubao Work can drive the browser and the local machine, run tasks across multiple apps, hand long-running work to a cloud computer that keeps going after the laptop is closed, and be picked up again from a phone. The user's company chat, in other words, is not just a context window. It is the operating system the agent moves through.

That coupling is also the audit story. ByteDance is positioning Doubao Work as a workplace tool, not a consumer toy. The release materials describe device enrollment, per-skill permission settings, quota controls, encryption, and an operation audit log, all inheriting Feishu's existing role and access system, with personal and corporate data kept on separate rails. Multiple specialised agents, for data, research, and design, can be assembled into a team on the same workspace, and the work they produce settles back into Feishu as reusable, editable, shareable enterprise knowledge. The pitch is that the agent does not just take from the company. It also leaves behind artifacts the company can keep.

The claims to watch are the ones ByteDance is making on its own behalf. The vendor says Doubao Work is the first Chinese office agent to pass both an "office-agent capability" assessment and a "cloud-side benchmark" test, a dual certification the company is using as a launch credential. The Xinhua wire that carried the news reads, on inspection, very close to the LeiFeng Wang version, which suggests the original signal is the company release rather than an independent regulator's verdict. Treat the certification as a sourced claim, not as a competitive fact.

Three things will tell readers whether this class of tool has crossed from demo to daily use. First, does reliability hold up on long, multi-step tasks, where one wrong click can derail a workflow and the user only notices at the end. Second, what does the audit log actually contain, and can a security team replay what the agent did and why, especially when several sub-agents cooperate. Third, what happens when the user wants the agent to do something the company has not blessed: the permission model and the data boundary will matter more than the model's benchmark score. ByteDance's bet is that binding tightly to Feishu answers all three. The next six months of user testing, in Chinese workplaces first and in any international Lark rollouts after, will say whether it does.