AI News Feed
Market watch
Products & Applications

Feishu and Doubao Work Launch Doubao Work Partner, a Team Agent for Group Chats

Feishu and Doubao Work released Doubao Work Partner, a team agent that joins group chats with its own organizational identity and permissions. The product is in early enterprise co-creation, and early testing shows it can handle documents, data, meetings, tasks and code changes within authorized Feishu contexts.

The release follows Anthropic's Claude Tag, a shared agent that can be added to Slack to join discussions, write code and analyze data. The report said Claude Code engineers estimated that 70% to 80% of their work was carried out through Tag.

Doubao Work Partner differs from personal office agents by having its own organizational identity and permissions. According to the report, it can read and reply in group chats it has joined, send messages to reachable people or groups, search public webpages, and use Feishu documents, meetings and tools within its authorization. Feishu's existing permission rules still apply. It can be added to different project groups, allowing product, engineering, operations and marketing discussions to connect through one agent.

In an early test by ifanr's APPSO, the agent was named Ai Jun and added to a work group. It answered questions, tracked tasks, processed Feishu documents, analyzed data, set monitoring and reminders, recorded to-dos, tracked owners and pushed work forward. It was available 24 hours a day. When told to help proactively on similar requests in the future, it recorded that preference in its memory.

The agent also performed hands-on work. The test asked it to find AI news every morning, assess whether each item was worth following and how timely it was, collect feedback and improve its topic selection. When a group member sent an X post link, it read the link, summarized the material and added the item to a topic library.

In a more involved project, the editorial team gave the agent a WeChat article formatting plug-in. After a group vote, it recorded a request to add a "download as Markdown" feature and classified it as a P2 candidate or new capability. It then unpacked the compressed file, modified the code, recompressed the plug-in into a ZIP file and sent it to the group in under ten minutes. When other colleagues found bugs, it debugged and fixed them. It also accepted screenshots and screen recordings as multimodal input, and helped collect feedback for a new name.

The report said the agent can attend meetings on a member's behalf using prior context from the project group, answer progress questions based on group discussions, and sync conclusions, to-dos and risks back to the group. It can remind owners of verbal commitments, push when no one responds and know whom to contact.

The report said the product is intended to shift office agents from individual tools to shared team capabilities. A task can continue around the same agent without people moving between chat windows, and the agent follows the project into different groups. This reduces repeated explanations, waiting, forwarding and alignment, according to the report.

The report said team agents need three conditions to work: complete organizational context, top-tier AI capability and enterprise trust. It said Feishu's all-in-one model keeps data, permissions and workflows together, while Slack relies on third-party apps for documents, spreadsheets and approvals. Feishu 8.0 will support calling agents in document comments and previewing or editing agent-generated Markdown and HTML files inside Feishu.

Doubao Work, launched in August, has added desktop Computer Use for Mac and Windows, browser automation, screen-recording skill learning, cross-device environment switching and multi-agent execution, according to the report. Doubao Work Partner has an independent organizational identity, and enterprises can configure which group chats, documents and tools it can access.

The product remains in early enterprise co-creation and is not fully available to all users. The report's testing was based on early access.