Doubao Work makes autonomy a permission ladder

Doubao Work frames agentic work around a project space where materials and results can accumulate. Reusable skills can accumulate there as well. According to woshipm, projects can link to local folders so task materials and generated results stay together, while its cloud drive can handle Doubao Work files alongside Feishu Drive files. It can also handle recently opened Feishu documents. Completed workflows can also become custom skills for later reuse.

That concentration of context makes an agent more useful, but it raises the stakes of what it may do alone.

geekpark reports that Doubao Work Partners has its own identity and permissions. It also has memory and can participate in group chats or meetings. The synthesis here is a permission ladder: begin with bounded access and drafting, then earn authority for consequential actions through review and clear separation of accounts. Memory and records should remain separate as well.

Build one bounded project space

A project becomes a bounded workspace when its source material stays connected to its working context. Its outputs remain connected as well. Doubao Work projects can link to local folders, placing task materials beside generated results, according to woshipm. That structure gives an agent a defined place to work rather than an invitation to range across a person's entire desktop.

The folder is the boundary.

Doubao Work's cloud drive extends that boundary across files generated in the product. It also covers Feishu Drive files and recently opened Feishu documents, woshipm reports. For repeatable work, this matters more than a broad promise of an AI coworker. A project can gather the relevant documents. It can retain the resulting drafts. It can also make the working trail legible to the people responsible for it. The agent's context is supplied by the project, not guessed from an unlimited collection of accounts and files.

Reusable skills make the arrangement durable. Doubao Work can automatically load local skills, and a completed workflow can be saved as a custom skill for later use, according to woshipm. That turns a successful sequence into a constrained operating pattern: start from the designated materials. Perform the established work, then return an output to the same project space. The value is consistency. This is especially true when a team wants a method reproduced without rebuilding instructions each time.

Doubao Work Partners adds another layer of separation. geekpark describes Partners as having independent identities. They also have permissions and memory. They can work in group chats and meetings. Partners can search across groups. They can convert a team's established methods into reusable capabilities. Those capabilities are useful only when attached to a clear project purpose. A linked folder and saved skill define what the partner is for; independent identity prevents the vague fiction that one general agent should inherit every context a team has.

Start with read access and drafts

The first rung of autonomy is read access. In Feishu 8.0, an Agent inherits its user's identity and data permissions, according to geekpark. It cannot inspect information the employee cannot view. That makes existing access rules the default boundary, rather than a separate AI permission model.

For Feishu users, Doubao Work's Enterprise Knowledge skill searches internal company documents through the user's own permissions, woshipm reports. A useful first task is therefore narrow: retrieve relevant material and summarize it. The employee can then identify gaps to resolve. The agent can draw on authorized group chats and documents. It can also use meetings, Bitable, or enterprise business systems through Feishu's shared account and permission system, within the limits of authorization.

Reading is useful because context precedes judgment.

The next rung is drafting. An agent with access to the relevant project material can turn a meeting record or internal document into a proposed update, then revise that draft after a user's edits. It is working from the employee's accessible context, not from a broad claim to company knowledge. That distinction reduces accidental exposure while making the draft more grounded.

Inherited permissions are also easier to administer. Doubao Work Partner applies Feishu organizational-structure rules, according to woshipm, allowing enterprise administrators to configure access scopes and set AI usage limits. They can also trace operation logs. The practical rule is to begin with work the agent can read. It can transform that work into drafts. Editing can follow when the user remains able to inspect the result and decide what becomes final.

Put sending and deletion behind review

Doubao Work exposes a broad choice of authority through its "Allow All" mode. It also offers "Confirm as Needed" mode, according to woshipm. The "Always Ask" mode is available separately. Those labels make permission a workflow decision rather than a fixed property of the agent. An operator can let a browser task proceed under close review, then widen authority only when the consequences are understood.

The critical boundary is between preparing an action and committing it.

That distinction matters because Doubao Work browser automation can open websites and read page content, woshipm reports. It can also download files or fill out forms. These actions can support useful progress without automatically giving an agent the right to publish a completed form or remove something already stored. "Confirm as Needed" is most valuable at that handoff: the agent assembles the work, while a person reviews the final irreversible step.

Mobile systems show why broad agent modes need limits; under SAEP, third-party apps can declare that AI screen automation is permitted. They can also restrict particular operations, according to the outlet. Apps using the consumer Doubao Mobile Assistant can refuse GUI operations altogether. They can block publishing or deletion. They can also block benefit collection. Doubao has also restricted financial payments and reward farming.

Google's Gemini multi-step tasks illustrate a similar checkpoint: they can arrange a ride-hailing trip with multiple stops. They can send progress notifications after moving into the background. They also require confirmation before a purchase, the outlet reports.

Authority should therefore be earned by task outcome; let an agent gather material and complete a draft, then require review when it sends or deletes. Review is also required when it collects benefits or pays.

Treat unattended GUI work as an exception

Desktop control can reach applications that have not built an AI interface. According to a report, after authorization, Doubao on the nubia M153 could simulate taps. It could also handle swipes and text input in third-party apps. That reach is useful, but it does not make a chain of app screens dependable. Carnegie Mellon University's iOSWorld benchmark used 26 self-developed iOS apps and 133 tasks; its best-performing model reached only a 37% success rate on cross-app tasks.

Treat scheduled GUI work as a narrow exception, reserved for actions whose effects are easy to reverse.

Doubao Work supports scheduled tasks that execute daily. It can also execute them weekly or monthly, according to woshipm. A sensible unattended job can prepare a recurring artifact from a stable source, then leave any ambiguous screen or consequential choice for a person. The interface makes this distinction practical: entering "/操作" in the chat box offers Computer Use and browser-operation options.

The handoff point should move earlier when an application contains a sensitive decision or an unclear permission state. A report says that, after the SAEP public-notice period, apps without a stated position may gradually be opened according to risk level, while apps that explicitly refuse will never be operated. That is a boundary, not a reliability guarantee. For work grounded in meeting records, woshipm says Doubao Work can draw action items from Feishu meeting minutes and smart summaries, then create personal Feishu tasks with deadlines based on meeting conclusions. Let the agent assemble that routine output; keep a person present when cross-app interpretation becomes uncertain.

Choose bounded workplace workflows; earn broader autonomy

• A task depends on a defined set of internal chats, documents, meetings, approvals, and business data. Use a workplace agent inside the organization's existing permission system rather than giving it a broad, separate data dump. Doubao Work shares Feishu's account and permission system, while Feishu 8.0 Agents inherit the user's identity and data permissions, preventing access to information the employee cannot view.
• A recurring task has stable inputs and a repeatable output, such as turning meeting conclusions into follow-up work. Build and reuse a narrow workflow. Doubao Work supports daily, weekly, or monthly scheduled tasks, can save completed workflows as custom skills, and can link projects to local folders so materials and generated results stay together.
• The agent needs to move from reading and drafting into actions that affect records, communications, or approvals. Start with staged permission rather than unattended execution. Doubao Work offers "Allow All", "Confirm as Needed", and "Always Ask" modes; Feishu administrators can trace Agent usage and control AI usage and high-consumption scenarios.
• The work can be completed through explicit workplace APIs and authorized product capabilities. Prefer those integrations over screen-driving automation. Feishu's CLI supports work such as operating Bitable, scheduling calendar tasks, accessing Drive and Minutes, initiating approvals, and drawing editable infographics in Whiteboard.
• A proposed workflow relies on GUI automation across third-party applications, especially for consequential actions. Treat reliability and authority as unproven until the task is tested and constrained. Carnegie Mellon University's iOSWorld benchmark reported a best cross-app-task success rate of only 37%; Doubao's mobile assistant also restricts high-risk scenarios including financial payments and reward farming, while SAEP lets apps reject GUI operations or restrict publishing, deletion, and benefit collection.

Keep memory, accounts, and logs separable

A project-linked local folder gives Doubao Work a visible boundary: according to woshipm, task materials and generated results can be stored together. That makes later inspection more practical, because the inputs and outputs belong in the same place rather than being scattered across personal drives. The folder is a useful evidence container, though it does not by itself prevent an agent from drawing on broader connected knowledge.

Memory needs its own boundary. An identity may be separate while its retained knowledge still travels.

That distinction matters for Doubao Work Partners. geekpark reports that Partners have independent identities. Their permissions and memory are also independent, while they can search information across groups and turn established team methods into reusable capabilities. woshipm further says a Partner can retain chat history and document contents. It can also retain meeting conclusions and project progress, then reuse that team memory across groups and projects. Teams should therefore inspect both the account scope and the memory scope: a separate project space is not sufficient if the agent's remembered context can cross into another group's work.

Administrator controls provide the second inspection point. According to woshipm, Doubao Work Partner follows Feishu organizational-structure rules for permissions, with configurable access scopes and AI usage limits. It also provides traceable operation logs. geekpark similarly says Feishu administrators can manage enterprise Agents centrally and trace usage. They can also control AI use in high-consumption scenarios. Logs make actions reviewable after the fact; scoped accounts limit what can be reached beforehand.

The same principle applies to connected suites. woshipm says WorkBuddy can directly use WeCom documents and spreadsheets through its CLI interface. It can also use email, calendars, and contacts. It also says Qwen Office can read and write DingTalk data. Those connections are valuable only when teams can identify the account and the permitted scope behind each action. They must also identify the recorded operation.

How workplace and mobile agents differ in context, permissions, and execution

Treat a workplace agent's available commands as a permission map, not a feature list. Qwen Office can read and write DingTalk data for group-chat management. It also supports meeting efficiency. It supports document collaboration. Process automation is also supported. It reaches 25 business modules and more than 1,000 commands, according to woshipm. Put the approval point in the request before granting write access.

Start with a dedicated folder and a draft output. WorkBuddy uses WeCom's CLI interface. It handles documents and spreadsheets. It can also access email. It can access calendars. It can access contacts. Keep execution gated until a person reviews the proposed change.

For readers outside China

• Availability: Doubao Work is downloadable from doubao.com/work, which identifies Mac and Windows devices and provides the corresponding version. It supports login with a Doubao account or a Feishu account. Availability outside China is not disclosed in sources. Doubao Work Partners is currently in targeted co-creation with enterprises and gradual opening.
• Pricing: Pricing is not disclosed in sources.
• Closest Western equivalents: Claude Tag on Slack, which launched as an independent participant in Slack channels with channel-level memory and permissions.; Google's Gemini multi-step tasks, which can arrange a ride-hailing trip with multiple stops, allow monitoring and takeover through notifications, and require confirmation before purchase.; A workplace automation layer combining collaboration-suite permissions, scheduled workflows, and browser or desktop Computer Use.
• Data residency: Doubao Work lets users choose whether a task runs locally or in the cloud, and projects can be linked to local folders. The sources do not say where cloud-run task data is stored, whether data can be processed outside China, or what data-residency commitments apply. For Feishu-linked use, access is described as operating within the user's authorization limits and existing account permissions.

Sources

• woshipm 豆包Agent最系统手把手教程，从入门到精通（附完整PDF） https://woshipm.com/ai/6459785.html
• geekpark 在飞书的上下文底座上，豆包开工了 https://geekpark.net/news/370460
• 36kr 豆包手机明日开卖，AI能帮你操作App了吗？ https://36kr.com/p/3984764117007111
• woshipm 豆包工作打出"团队Agent"牌，Workbuddy、千问会跟吗？ https://woshipm.com/it/6465348.html