ByteDance Ships Doubao Work, Its First Agent Product After Merging Feishu and Doubao
What This Change Actually Does
The core change in Doubao Work revolves around two dimensions: "independence" and "integration."
Previously, the "work tasks" feature inside Doubao was essentially a lightweight task-execution entry point. Users had to manually configure tools and guide the Agent step by step to complete work. Spinning it off into a full product means ByteDance is upgrading the Agent from a "feature module" to a "product form" — clearer positioning and more explicit capability boundaries.
Integration with Feishu is the key differentiator. Doubao Work supports one-click login with a Feishu account, fully inheriting enterprise AI quotas, documents, multi-dimensional tables, knowledge bases, chats, email, and the Feishu permission system. This means the Agent can directly read enterprise context without requiring users to re-enter background information, and the accuracy of task decomposition and tool invocation will improve significantly.
Another noteworthy change is computer-control capability. Doubao Work can package fixed workflows into reusable Skills, which has a direct substitution effect for repetitive, rule-explicit work scenarios (such as data organization, report generation, and meeting minutes).
##What This Change Actually Does The core
Mechanism: Why Feishu + Doubao
The underlying logic of this integration is a three-layer architecture of "foundation + model + scenario."
Feishu provides enterprise context and the permission system, Doubao provides model capability and Agent orchestration, and Volcano Engine provides the underlying compute support. The three are shifting from relatively independent business units into a coordinated system.
From a technical implementation perspective, Doubao Work adopts a classic three-column layout: manage tasks on the left, view execution progress in the middle, and preview and edit results on the right. Users only need to state the goal, and the Agent will automatically decompose tasks, invoke tools, and continuously advance complex workflows.
This design is similar to products like Codex and Claude Cowork, but Doubao Work's differentiation lies in the complete inheritance of the Feishu permission system. Enterprise data does not leave the Feishu ecosystem, and the Agent executes tasks within permission boundaries, making it more attractive to data-security-sensitive enterprises.
##Mechanism: Why Feishu + Doubao This time
Who Will Feel the Pain First
Feishu CLI users are the first affected group.
Previously, heavy Feishu users relied on the Feishu CLI paired with Agent tools like Claude Code and Codex to execute tasks within Feishu. The Feishu CLI required frequent updates and had a high onboarding barrier, making it difficult for ordinary users to use proficiently. After the release of Doubao Work, these users can use the native product directly without additional CLI configuration.
Competing products WorkBuddy and Qianwen Office face direct competition.
According to Analysys' "China Office Agent Platform Market Insight Q2 2026" report released in July 2026, Tencent WorkBuddy had 20.97 million visits in June alone, exceeding the combined total of ByteDance's TRAE IDE (domestic version) and Alibaba's QoderWork, maintaining the market-leading position for two consecutive quarters. The release of Doubao Work will change this landscape.
##Who Will Feel the Pain First Feishu CLI users are the first
Actionable Steps
For enterprises already using Feishu, it is recommended to evaluate which scenarios in existing workflows are suitable for migration to Doubao Work. Focus on observing the substitution effect of Skill packaging capability on repetitive work, especially high-frequency scenarios like data organization, report generation, and meeting minutes.
For users of WorkBuddy or Qianwen Office, compare the differences in three dimensions — Feishu permission inheritance, computer control, and Skill packaging — to judge whether switching is worthwhile.
This conclusion holds within the Feishu ecosystem; beyond that scope, competitive advantages need to be re-evaluated.
##Actionable Steps For enterprises already using Feishu
Where the Mechanism Actually Gets Stuck
Permission Inheritance: Not as Simple as "Integration"
Doubao Work claims to "fully inherit the Feishu permission system," but behind this product rhetoric lies the cost of reconstructing the permission model. Feishu's permission system includes three-layer control at the space level, document level, and field level, and supports dynamic roles (such as "department admin" and "project observer"). For an Agent to truly inherit this system, mapping is needed at two levels: first, the identity authentication layer — the binding relationship between the Doubao account and the Feishu enterprise ID; second, the data access layer — the permission verification logic when the Agent reads documents.
QbitAI's hands-on testing points out that Doubao Work supports one-click login with a Feishu account, but the completeness of permission inheritance depends on the openness scope of Feishu APIs. Feishu open platform documentation shows that its APIs have explicit restrictions on permission applications for Agent-type applications, especially when it involves cross-department data reading, which requires manual approval by administrators. This means "full inheritance" will have gray areas in actual implementation — not a product-level issue, but a boundary issue of enterprise IT governance.
Context Window: The Real Capacity for Enterprise Data
Doubao Work's core capability is to "autonomously decompose tasks around user goals, invoke tools, and continuously advance complex workflows." What underpins this is the large model's context window. Doubao's daily Token invocation volume exceeds 180 trillion, but the context window for enterprise-grade Agents is not infinite.
Taking the Feishu knowledge base as an example, a mid-sized enterprise's knowledge base may contain tens of thousands of documents, with a total word count reaching hundreds of millions. For the Agent to "understand" this content, it needs to compress it into the context window. Current mainstream large models have context windows between 32K and 128K Tokens, which translates to roughly 20,000 to 80,000 Chinese characters. This means the Agent cannot read all enterprise knowledge at once and must rely on Retrieval-Augmented Generation (RAG) technology.
The bottleneck of RAG lies in retrieval precision. Vector retrieval in the Feishu knowledge base depends on document metadata tags and embedding quality. If enterprise knowledge base documents have non-standard naming and missing tags, retrieval recall rates will drop significantly. This is not a problem unique to Doubao Work, but a common challenge faced by all enterprise-grade Agents.
Task Orchestration: Where Are the Boundaries of Skill Packaging
Doubao Work supports packaging fixed workflows into reusable Skills. This sounds like a universal solution, but the packaging boundaries of Skills determine the upper limit of the Agent's capability.
Taking an MCN talent agency team as an example, if the Agent needs to handle the task of "contract review for signing a blogger," the Skill needs to call Feishu documents to read the contract, call AI for clause analysis, and call the email system to send review opinions. This process involves three systems and four types of tool invocations, and each step may fail (e.g., insufficient document permissions, AI analysis timeout, email sending blocked).
The difficulty of task orchestration lies in error handling and the timing of Human-in-the-Loop (HITL) intervention. Redis documentation on AI Agent Pipelines points out that reliable Agents need an "evaluator-optimizer loop" — continuously evaluating results and adjusting strategies during execution. Whether Doubao Work's Skill mechanism supports this kind of loop is not explicitly stated in currently public materials. If Skills are static processes that require human intervention once an anomaly is encountered, then the capability to "autonomously advance complex workflows" will be greatly diminished.
Who Will Feel the Pain First: IT Department or Business Department
The release of Doubao Work will first impact the IT department. Permission inheritance, context management, and task orchestration — these three mechanism bottlenecks all require the IT department to adapt internally within the enterprise. The integration of Feishu and Volcano Engine means ByteDance is packaging and selling "foundation compute + large model + front-end application." After procurement, enterprises need to face the reconnection of internal systems.
Business departments are indirect beneficiaries. The Agent can handle repetitive work such as writing proposals, organizing meeting minutes, and drafting reports, but the premise is that business departments are willing to open data to the Agent. If enterprise data sensitivity is high (such as in finance and healthcare industries), business departments may refuse to let the Agent access core data, limiting the Agent's capabilities.
Actionable Steps: What to Do Now
For enterprises considering introducing Doubao Work, the following three steps are recommended:
First, evaluate the openness level of the Feishu permission system. Contact the Feishu customer success team to confirm the API scope accessible to Agent-type applications, especially the approval process for cross-department data reading.
Second, review the metadata quality of the enterprise knowledge base. If document naming is chaotic and tags are missing, it is recommended to do data governance first before introducing the Agent. Otherwise, retrieval recall rates will directly affect the Agent's output quality.
Third, design the exception handling mechanism for Skills. When packaging fixed workflows, clarify the failure retry strategy and human intervention nodes for each step to avoid the Agent getting "stuck" during execution.
The release of Doubao Work marks ByteDance's official entry into the enterprise Agent arena. But the existence of mechanism bottlenecks means this is not an "out-of-the-box" solution. Enterprises need to adapt at three levels — permissions, data, and processes — to realize the Agent's true value.
##Where the Mechanism Actually Gets Stuck###Permission Inheritance
Judgments That Can Be Made Today
Choice Paths for Enterprises of Different Sizes
The release of Doubao Work shifts enterprise Agent implementation from "whether to do it" to "how to do it." Enterprises of different sizes have different choice paths.
Large enterprises (more than 500 employees): It is recommended to first establish an Agent governance committee to clarify permission boundaries and usage norms. Doubao Work fully inherits the Feishu permission system, which means incorrect permission configuration could lead to unauthorized access. It is recommended that the IT department first review the existing Feishu permission system, identify high-risk operations (such as financial approvals, customer data exports), and list them as boundaries that the Agent cannot directly operate. Pilot in low-risk scenarios (meeting minutes organization, proposal drafting) first, verify accuracy, and then gradually expand.
Mid-sized enterprises (100-500 employees): Can directly pilot in scenarios such as marketing, customer service, and operations. The inputs and outputs in these scenarios are relatively standardized, and the impact of Agent errors is controllable. It is recommended to first use Doubao Work's Skill function to package high-frequency repetitive processes (such as weekly report generation, competitive research) into reusable templates to lower the usage barrier.
Small enterprises (fewer than 100 employees): If already deeply using Feishu, Doubao Work is the most direct choice. Migration cost is lowest, and documents, spreadsheets, and knowledge bases within the Feishu ecosystem can be directly invoked. If still using other collaboration tools (such as DingTalk, WeCom), migration costs need to be evaluated. Doubao Work currently supports connectors for WeCom and DingTalk, but full functionality is still primarily centered on Feishu.
Three Boundaries That Must Be Clarified
Doubao Work is not a panacea. The following three boundaries must be clarified.
First, the Agent is not suitable for handling sensitive tasks involving financial decisions or customer privacy. Although Doubao Work inherits the Feishu permission system, permission inheritance does not equal permission understanding. The Agent may operate within the permission scope but cannot judge the business rationality of the operation. Scenarios involving fund outflows, contract signing, and customer personal information processing must retain a manual review step.
Second, the accuracy of complex tasks still depends on human intervention. Doubao Work can autonomously decompose tasks and invoke tools, but the error accumulation effect in multi-step tasks is significant. For a 15-step process with 95% accuracy per step, the final success rate is only 46%. It is recommended to break complex tasks into multiple independent Skills, with human confirmation after each Skill is completed before proceeding to the next step.
Third, Skill packaging has applicable boundaries. Doubao Work supports packaging fixed workflows into Skills, but scenarios with frequent process changes are not suitable for packaging. Once a Skill is published, modification requires reconfiguration and testing. It is recommended to first identify high-frequency, stable processes (such as monthly report generation, contract template filling) before considering packaging.
Next-Step Action Checklist
If deciding to try Doubao Work, the following three things can start today:
First, apply for a test account. Doubao Work is already available for download; log in with a Feishu account. It is recommended to first let 2-3 core users familiar with Feishu operations try it, collect feedback, and then promote it.
Second, identify three high-frequency scenarios. Choose the three most repetitive, rule-clearest tasks within the team (such as meeting minutes organization, competitive information collection, weekly report generation), attempt to complete them with Doubao Work, and record accuracy and time spent.
Third, establish usage norms. Clarify which scenarios can use the Agent and which must be handled manually. The norms don't need to be complex — one page is enough — but team members must be informed.
The release of Doubao Work marks enterprise Agents moving from "usable" to "easy to use." But easy to use does not mean mindless use. Permission boundaries, accuracy boundaries, and process boundaries — these three lines of defense must be maintained.
References
[1] The first Agent product after the merger of Feishu and Doubao, Doubao Work, it's here._Data_Spreadsheet_Code. https://www.sohu.com/a/1067227582_121675819 [2] The first Agent product after the merger of Feishu and Doubao, Doubao Work, it's here_Tencent News. https://view.inews.qq.com/a/20260825A04UNB00?uid%5B0%5D=465365288&uid%5B1%5D=4653652883 [3] The first Agent product after the merger of Feishu and Doubao, Doubao Work, it's here_Tencent News. https://view.inews.qq.com/a/20260825A04UNB00 [4] The first Agent product after the merger of Feishu and Doubao, Doubao Work, it's here. https://zhuanlan.zhihu.com/p/2075583054949753252 [5] Doubao Work officially released, deeply integrated with Feishu to build enterprise-grade Agent _ Eastmoney. https://finance.eastmoney.com/a/202608253852692063.html [6] ByteDance releases Doubao Work, deeply integrated with Feishu to build enterprise-grade Agent. https://www.stdaily.com/web/gdxw/2026-08/25/content_569505.html [7] In-depth hands-on 'Doubao Work' + Feishu: currently the closest answer to the endgame of enterprise Agents – QbitAI. https://www.qbitai.com/2026/08/[REDACTED].html [8] ByteDance major adjustment: Feishu products merged with Doubao, commercialization merged with Volcano Engine_Sina Finance_Sina. https://finance.sina.com.cn/stock/t/2026-07-30/doc-inikqcsm1292038.shtml