The next-generation productivity is right within Doubao Work + Feishu.

In 2026, airport billboards, elevator screens, and WeChat Moments feeds — overwhelming advertisements are all promoting one same offering: AI that helps you work. 

If you line up the feature lists of products from different vendors, you will find they are looking increasingly identical. This means the era of building competitive barriers through a long feature checklist has come to an end.

On August 25, "Doubao Work" is officially launched, a brand new Agent product and brand, which also highlights a core feature: deep native integration with Feishu.

This seemingly trivial announcement pushes the competition into the second half: the first half was a race to build more capable Agents, while the second half is a race to deliver tangible improvements to organizational productivity.

Agents Are Getting More Powerful, But Enterprises Are Not Speeding Up

Let's first look at what happened in the first half of the Agent race.

Since the launch of OpenClaw, the gaps between large language models have been narrowing, and the engineering capabilities required to turn models into Agents are also being rapidly standardized. Memory, Skills, MCP, tool calling, Sandbox, Computer Use, long-task execution... The Agent Harness that used to require a dedicated team to build from scratch is now increasingly provided directly by model vendors and open source frameworks.

In April this year, OpenAI updated the Agents SDK to incorporate capabilities including memory, sandbox, file system tools, MCP and Skills into a standardized Harness. Lang Chain later even summarized Agent with a simple formula: Agent=Model+Harness. The increasingly mature SDKs and open source Harness allow developers to swap models, connect tools, and reuse proven mature Agent architectures.

This reshapes the competition logic of Agent products — feature lists are no longer a competitive barrier.

After a new function rolls out on one product, the time for other products to follow up is shrinking; once a Harness is verified to be effective, it may soon become reusable infrastructure for the entire industry.

This drives the rapid growth of Agent capabilities: they can process longer files, and handle more types of tasks.

However, the speed at which enterprises convert AI capabilities into real organizational productivity has not kept pace synchronously.

This mismatch has already been reflected in third-party research. Deloitte's 2026 survey of 3235 enterprise executives across 24 countries shows that the coverage rate of AI tools provided to employees has risen from less than 40% to 60%, but only 34% of enterprises have started to use AI to deeply transform their products, core processes or business models. Another 37% of enterprises still stay at relatively superficial application levels, making almost no changes to their existing business workflows.

This gap, when reflected on every employee in the enterprise, is very concrete.

For example, an employee wants the Agent to "review a project" or "find out why the project was delayed". Even though large language models are highly competent at summarization and reasoning, the real obstacles often emerge before the AI even starts working: what new requirements the client just modified, what conclusions were reached in last week's meeting, which version is the latest solution, who is responsible for the progress, and at which approval step the process is stuck — all this information that determines judgment will not naturally appear in the Agent's dialog box.

As a result, employees have to first scroll through group chats, find meeting minutes, download spreadsheets and documents, organize information scattered across different locations, and then re-explain all the context to the AI.

The Agent seems to be working for people, but people have to prepare work for the Agent first.

Work in enterprises is never a collection of isolated tasks. A requirement may start from discussions in group chats, reach consensus in meetings, be deposited in documents and spreadsheets, then enter projects, task lists and approval flows, and keep circulating between different departments and responsible persons. Its antecedents and consequences are scattered in knowledge assets, data, workflows, permissions, organizational relationships and historical decisions.

Capabilities determine whether an Agent can perform tasks, while context determines whether an Agent can perform tasks well.

At present, the former is no longer a bottleneck, while the latter is.

Doubao Work: Finding Its "Workstation" Inside Feishu

To solve this bottleneck, the most critical move made by Doubao Work is account-level deep integration with Feishu.

After logging in with a Feishu account, the Agent can directly inherit the user's work context in Feishu — it can access chat records, documents, meeting minutes, schedules and other information within the scope of the user's existing permissions. It does not require employees to manually move scattered information over, but directly searches for information, understands relationships and continues to take actions where the work actually happens.

The significance of this move depends on what Feishu contains in the first place.

In most enterprises, data is fragmented. When an Agent wants to complete a cross-system task, it has to connect each system one by one, migrate data, and then process identity and permission mapping between different systems. Every additional system adds an extra layer of friction.

But Feishu natively aggregates all these assets: chats, documents, meetings, knowledge bases, multidimensional spreadsheets, projects and approvals all run under the same set of organizational identities, permissions and collaboration systems.

More importantly, Feishu has been building infrastructure for AI invocation over the past few years. Meetings are automatically saved as searchable, quotable Smart Records, and documents, knowledge bases, group messages and multidimensional spreadsheets have all been adapted for AI invocation scenarios.

The data is not just piled up there, but has been organized into a format that Agents can understand and use directly.

Among mainstream domestic collaboration platforms, Feishu has the highest degree of unified organizational information and the most complete adaptation for Agents, making it closer to an AI-native organizational operating system. This is why under the "Lobster" Agent development wave, Feishu has been naturally regarded by developers as the core execution and collaboration entry point for Agents.

When Agents actually run on such a solid foundation, the efficiency improvement brought to work is immediate.

In the past, the actual progress of a cross-departmental project was usually scattered in group chats, meeting minutes, task lists and approval flows. To confirm whether the project can be launched on schedule, employees often need to align status with every stakeholder one by one. But now, you only need to ask Doubao Work "Can we still launch this on X date", and the Agent can follow the traces left in Feishu, search for information automatically, and pull out the full progress and risk map. When encountering key nodes such as GPU expansion, resource application, and version release, it can also trace who submitted the application and when, and what impact approval or rejection will have on the project.

Employees no longer need to spend half an hour supplementing background information for the AI. The Agent starts to work like a colleague who has participated in the whole project from the very beginning — it searches for information on its own, sorts out all antecedents and consequences, and then gives its judgment.

But merely "knowing what is happening in the company" is not enough. A real digital employee also needs to be able to push tasks forward, and leave the results inside the organization.

Doubao Work can process documents, spreadsheets and PPTs, build web pages and applications, generate images and videos, and operate browsers and local computers. But the real difference does not lie in these functions themselves, but in where the outputs go.

If an AI-generated report stays only in a personal dialog box, colleagues cannot view or edit it, and the user still needs to re-explain all the background next time they need it — this one-off efficiency gain will disappear when the next work starts.

But after connecting to Feishu, the content generated by Doubao Work can be saved back to Feishu, shared, commented on and edited by colleagues; mature working methods can be precipitated into organization-shared skills that can be directly invoked next time.

This forms a complete closed loop: understanding the organization, executing tasks, joining collaboration, receiving feedback, and precipitating into new organizational experience. The more the Agent works, the more its outputs are converted into enterprise assets, which in turn become reusable context for subsequent work.

A tool can perform tasks, but only a "digital employee" that knows what is happening in the company, can push tasks forward independently, and leave results for the team is a real Agent.

Agent ROI: From Individual Efficiency Improvement to Organizational Efficiency Growth

After Agents enter the organization, the criteria for enterprises to measure the value of AI will also change accordingly.

Compared with how many minutes an individual saves from using AI, a more important question becomes whether AI can reduce the communication, collaboration and process costs of the entire organization.

BCG's global "AI at Work" survey released in June this year gives a striking figure: 42% of frontline employees who use AI frequently can save at least 8 hours per week. But BCG immediately pointed out that most enterprises have not yet figured out how to convert these saved hours into organizational value.

After all, one person working faster does not equal the whole team working faster.

The majority of real efficiency losses in enterprises do not exist within individual tasks, but between tasks, and between people: searching for materials, asking for progress, synchronizing information, confirming responsible persons, cross-departmental communication, and waiting for approvals. None of these actions are complicated when viewed separately, but they together constitute the largest hidden cost in the daily operation of organizations.

AI can compress the time spent making a PPT from two hours to half an hour, but if the subsequent review, modification and follow-up of this PPT still follow the old workflow, the organizational-level efficiency will not be improved at all.

This means the value of Agent is not to help people finish a piece of material, but to take over the whole series of connecting actions after the material is generated: to search, to follow up, to synchronize, and to push the process forward.

This requires it to truly integrate into the business scenarios of the organization: client data, operation metrics, meeting content, and approval workflows.

As a result, permission management becomes an unavoidable prerequisite.

Client information that an employee has no right to view should not be accessible to the Agent either; operations that an employee is not authorized to perform cannot be bypassed by the AI.

Doubao Work inherits Feishu's existing identity and permission system, so the Agent can only access data and operate tools within the permission scope of the current user; at the same time, Doubao Work has built a strict full-link Agent security protection system covering device access, permission configuration, quota control, data encryption and operation auditing. At present, it has become one of the first batch of office Agents in China to pass the dual certifications of "Office Agent Capability" and "Cloud Benchmark Test" from the China Academy of Information and Communications Technology.

This is not a more powerful Agent, but a design that allows the Agent to run as a member of the organization within the existing security governance system.

In the past, enterprise software managed what people can do; after Agents are introduced, it also needs to manage what AI can do on behalf of people — "permission" is becoming a key dividing line between organizational-level Agents and personal AI assistants.

From this perspective, the evaluation system for this round of Agent competition has also changed.

In the first half of AI office competition, all players competed to provide more powerful AI for individual users: smarter models, higher generation quality, and more completed tasks. These capabilities are converging, and the gaps between different products are getting smaller and smaller.

In the second half, what can truly differentiate products is whether the Agent can safely enter the organization, understand ongoing work, participate in real workflows, and turn the results of a single task into experience that the entire organization can reuse.

Therefore, the truly remarkable part of Doubao Work is not that there is another Agent that can make PPTs, but that ByteDance has tightly integrated its full-bet Agent product with the most AI-Native organizational infrastructure.

The competition of Agents is also shifting from individual productivity to organizational productivity.