ByteDance's AI office assistant, DouBao Work, released this week's update on September 23: After the model update and the release of "DouBao Work Partner," task mode now adds "Goal" mode and "Plan" mode, and also launches task queue, Markdown file editing, and dark mode. These features are now fully rolled out on desktop clients, and you can experience them by updating to the latest version.

QQ20260923-135202.jpg

Goal mode is designed for complex tasks with clear requirements: After users clearly state the acceptance goals, they can start execution. Once the task is completed, DouBao Work will check each goal one by one. If any item does not meet the standard, it will automatically rework and continue adjusting until it meets the standards.

Plan mode is suitable for complex tasks with many steps and high resource consumption: DouBao Work will first generate a Markdown format implementation plan that can be reviewed and edited, allowing users to calibrate and modify it. The execution will only start after aligning with the work requirements and confirming. This makes the execution process more controllable and reduces resource and time consumption. In this mode, online editing, viewing, and sharing of .md files are also supported. You can also issue instructions for DouBao Work to make direct modifications. Both modes can be activated by typing "/" in the chat bar or clicking "+" at the bottom left.

The newly launched task queue feature allows users to send multiple new instructions continuously without interrupting the current task execution. New instructions are queued and executed automatically in order once the current task is completed. The newly added dark mode offers a more immersive experience, and its appearance can automatically switch according to the computer's appearance, and it also supports individual settings.

From "generating results" to "goal verification + plan confirmation," DouBao Work is bringing the acceptance loop and execution forward into AI office products. Agent collaboration is moving from one-click execution to controlled processes, indicating a clear direction for the productization of AI office assistants.