Recently, ByteDance-affiliated enterprise productivity products have released two updates: Volcano Engine has upgraded the Doubao large model Doubao-Seed-2.1-pro to version 0915, with full API launch on Volcano Ark and availability in Doubao Work; Feishu launched version 8.0 at the 2026 Future Unlimited Conference and Doubao Work Kick-off Event, and simultaneously introduced the Team Intelligent Agent "Doubao Work Partner," bringing AI from personal assistants into group chats, documents, meetings, and business processes.
On the model side, version 0915 focuses on "production-level delivery." In terms of Agent tasks, the new model enhances evidence tracing, authoritative source retrieval, timeliness judgment, and data verification, reducing hallucinations in scenarios such as financial research and office automation that require multiple rounds of tool calls, online research, and long report outputs, making complex long-term tasks more stable. In an official demonstration, the model can coordinate over 500 sub-Agents to search more than 1,000 web pages, cross-compare information on financial reports, production capacity, fleets, and hiring, and generate traceable due diligence reports by combining satellite imagery and maritime data as external evidence.
Multi-modal programming is another main line of this update. The new model better understands cross-file, long-term issues in large code repositories, can clarify module relationships, locate root causes, and run verification. In a self-repair test of the open-source game Luanti with about 387,000 lines of code and 1,000 real historical Issues, multiple sub-Agents worked in parallel for 36 hours, achieving merge standards for 83% of the issues; in an old ERP scenario, just with a screen recording and a few sketches, it could generate mobile pages for purchase orders, inventory, and stock based on a 280,000-line system. Multi-modal understanding also covers 3D and professional graphics, and with just four design drawings, it can recreate an interactive courtyard scene with water bodies, vegetation, camera angles, and seasonal conditions.
In terms of cost and access, version 0915 also optimizes Token efficiency, with significantly reduced token consumption for image and video reasoning compared to the previous generation. APIs on Volcano Ark provide two entry points: the locked version Doubao-Seed-2.1-pro-0915 and the automatically updating Doubao-Seed-Evolving. After the Doubao Work update, users can choose "Doubao 2.1 Pro (New 0915 Version)," and TRAE has also been integrated.
On the collaboration scenario side, Feishu 8.0 restructures through intelligent agent collaboration: users can add "Doubao Work Partner" to group chats as they would invite colleagues, and multiple intelligent agents can cooperate within authorized ranges, calling Feishu documents, multi-dimensional tables, meetings, calendars, cloud drives, and enterprise business systems to perform information synchronization, document review, weekly report organization, and code analysis. It has its own identity, permissions, and memory, remembering team experiences from group chats, documents, and meetings, and actively stepping in according to project progress. Enterprises can uniformly configure capabilities, data scope, and usage, managing the permissions, audit, and security boundaries of Agents following the logic of managing people.
Feishu stated that the Doubao Work Partner is currently in a targeted co-creation phase with enterprises and will be gradually opened up. Its goal is not to replace single-consultation office AI but to make AI a "virtual member" that resides permanently in the team, shifting from "each person independently calling a model" to "the whole team sharing one AI colleague who follows up, hands over, and is traceable."
Industry opinions suggest that the signal of this update lies in the shift of enterprise AI from "single-point competition of model capabilities" to a full chain: Doubao handles models and task execution, Feishu carries organizational context and permissions, and Volcano Ark provides APIs, computing power, and development infrastructure. If it can successfully close the loop in real projects for verification, programming, documentation, and meetings, enterprise office AI agents are expected to move from demonstration tools into routine production.
