How Doubao Evolved from a Chatbot into a Desktop AI Agent

Doubao’s PC client is no longer best understood as just a chatbot. The more useful description is a desktop AI agent: a system that can interpret a goal, select tools and carry out multiple steps across browsers, connected services and computer interfaces.

That makes “Chinese ChatGPT-plus-Codex” a reasonable product analogy. But it is still an analogy, not proof that Doubao is the best domestic replacement for Codex. The available evidence shows a significant expansion in execution capabilities, while comparative data on reliability, security, integrations and developer workflows remains limited.

The key change: from generating answers to completing workflows

The clearest evidence comes from Doubao’s Windows Work Task Mode and virtual desktop. After authorization, the agent can view the computer screen, move the mouse, click buttons and type on the keyboard. The user can observe the process, pause it or take control. 1

That changes the basic interaction model. A conventional chatbot might explain how to complete a task or draft the required content. A computer-use agent attempts to continue into the applications themselves: opening software, entering information, moving data between tools and producing an output.

The broader PC-client architecture has also been described as combining authenticated browser access, GUI computer operation, remote execution from a phone to a PC, selectable cloud computers, connectors and reusable Skills. 27 Together, these features resemble the direction of an AI assistant that combines ChatGPT-style conversation with Codex-style task execution.

Why the ChatGPT-plus-Codex comparison makes sense

The comparison has two layers:

• ChatGPT-like assistance: natural-language interaction, research, writing, document work, multimodal creation and task planning.
• Codex-like execution: multi-step action across browsers, connected services, desktop applications and, when necessary, the graphical user interface. 27

Doubao’s product logic is therefore broader than “an AI that can click.” The agent can serve as an interface for tools, workflows and specialized capabilities. ByteDance documentation references web-search plugins, Agent Skills and MCP, pointing toward a tool-oriented system rather than a single conversational window. 17

The distinction matters because GUI automation is only one layer. Connectors and structured tool calls can provide more predictable access to services, while GUI control is useful when no suitable integration exists.

Skills turn individual workflows into reusable processes

A Skill is best understood as a packaged workflow: a repeatable sequence that helps the agent perform a particular kind of task. In practice, the most valuable Skills are likely to be those with:

1. Clear inputs;
2. Repeatable steps;
3. An output that can be checked; and
4. A defined point for human approval.

This is a more practical approach than asking an agent to handle an enormous, ambiguous objective in one attempt. Start with a small process, inspect the result, correct the failure points and only then consider reusing or scheduling it.

Reporting around Doubao Work describes office production, research, scheduled analysis and team-shared workflows, including the possibility of returning documents, spreadsheets, presentations and web results to Feishu for continued review. That account comes from a user-generated social-media source, so these integration and availability claims should be treated as lower-confidence product reporting rather than independently verified capability. 32

The underlying principle is durable: AI becomes more useful when people deliberately identify repetitive, rule-based work and convert successful sequences into supervised, reusable processes.

Where the demonstrations fit—and where they do not

Examples such as comparing products on JD.com, creating a Tencent Surveys questionnaire, arranging SF Express shipping, completing invoices, comparing API prices, analysing YouTube titles and running scheduled workflows all illustrate the same proposed pattern:

Describe the desired outcome, let the agent execute routine steps, then review the result.

However, the supplied evidence does not independently verify each of those specific demonstrations as a generally available or reliable capability. A product demo may depend on a particular account state, website layout, permission setting or manually completed step.

The better-supported claim is narrower: Doubao’s virtual desktop demonstrates computer-use automation in which the agent can operate a local Windows environment under user authorization and supervision. 1

That is important, but it is not the same as proving reliable automation across every website, payment flow, enterprise system or edge case.

Seedance 2.5 adds a creative-agent dimension

Doubao’s agent strategy is not limited to office software. Seedance 2.5 was introduced with improvements in long-form narrative creation, multimodal references and editing, and was made available through Doubao Professional. 18

The model supports native generation of a single 30-second video, while Doubao’s creative-video Skill can use an agent workflow to break down a broader brief and coordinate image and video generation. 26 This gives Doubao a distinctive product shape: the same assistant can potentially move from a natural-language brief to creative planning, reference assets and a finished video workflow.

That does not establish that the resulting videos are superior in every professional setting. It does show how Skills can act as a layer between a user’s broad goal and several specialized models or tools.

A safer automation hierarchy

The safest way to use a desktop agent is to choose the least fragile automation method that can complete the job:

1. Use an official connector or API-backed integration first. Structured integrations generally make permissions, inputs and outputs easier to understand and audit.
2. Use CLI or MCP only when the integration is approved and the data scope is clear. These tools can be powerful, but they should not receive broader access than the task requires.
3. Use GUI computer control as a fallback. It is valuable for applications without integrations, but screen-based actions can be brittle and may trigger the wrong control or an irreversible action. Doubao’s pause-and-takeover design is therefore an important part of the workflow, not a minor interface detail. 1
4. Require approval before consequential actions. Keep a person in the loop for logins, payments, account changes, external publishing, government submissions and final approvals.

Avoid unsupervised operation on risk-sensitive services, including financial platforms, payment systems, accounts that can publish publicly, portals containing sensitive personal information and services such as Xiaohongshu where an erroneous post or account action could be costly.

The practical rule is simple: let the agent prepare and navigate, but make the human confirm anything difficult to reverse.

Market scale gives Doubao a distribution advantage

Doubao’s reported scale helps explain why ByteDance is pushing the product from conversation toward work execution. One report puts Doubao at 382 million monthly active users, while a Beijing government report also cites more than 382 million monthly active users. 29

For comparison, Reuters reported that ChatGPT reached an estimated 1 billion global monthly active app users in May 2026, based on Sensor Tower data. 33 These figures are not directly equivalent: they differ in geography, measurement context and product scope. They do not by themselves establish model quality, enterprise readiness or agent reliability.

Doubao’s opportunity is therefore less about matching ChatGPT on a single global user metric and more about converting an enormous China-focused user base into habitual use of local work tools, services and creative workflows.

The verdict: promising alternative, unproven winner

The strongest defensible conclusion is that Doubao has become a compelling China-oriented Codex-style alternative. Its case rests on the combination of:

• Windows virtual-desktop computer control; 1
• Browser, remote-execution and connector capabilities reported for the PC client; 27
• Skills, web-search plugins and MCP support; 17
• A large reported user base; 29
• Native access to ByteDance’s creative-video ecosystem, including Seedance 2.5. 1826

What the evidence does not establish is that Doubao is categorically the best domestic replacement for Codex. That claim would require controlled comparisons of task-completion rates, failure recovery, security controls, integration coverage, pricing, developer tooling and production reliability.

For now, Doubao’s most important evolution is conceptual: it is trying to make AI useful not only for producing an answer, but for moving work forward. The winners in this category will be the products that pair broad capability with reliable integrations, transparent permissions and disciplined human supervision.