Six domestic AI giants are vying in the Agent arena. Can they even catch a glimpse of Claude Code's back?

In the past six months, Agent products represented by Claude Code/Cowork and Codex have undoubtedly been the clearest main line in the entire AI industry.

On June 30th, Anthropic quietly launched Claude Science, an AI workbench for scientists.

Image source: Anthropic

Anthropic's thinking is becoming clearer and clearer. First, let the Agent take over code work (Claude Code), then let it handle various office tasks (Claude Cowork), and finally enter more complex professional scenarios such as scientific research (Claude Science).

In fact, Anthropic noticed that non - technical teams such as marketing and data teams started to use Claude Code bypassing the chat interface, which led to the creation of Claude Cowork with a graphical interface "encapsulation".

Besides the continuous iteration of the underlying model, many people may not be aware of the catch - up of domestic AI in Agent products. Alibaba, Tencent, ByteDance, Kimi, MiniMax, and Zhipu are all working on Agents. Each company basically follows Anthropic in distinguishing two types of products: Code and Work, but with different focuses.

So, before the model capabilities of domestic AI can comprehensively surpass those of top overseas products, can domestic AI rely on Agent products that are cheaper, more open, and closer to the local workflow to lead more domestic users to the work mode of "I command the Agent to work"? Lei Technology AGI (ID: leikejiagi) conducts a review to try to answer this question.

Tencent WorkBuddy: Aiming to be the office entry for more people

Tencent WorkBuddy may be the most well - known Agent product in China. After all, many people may have seen its advertisement on WeChat and installed and experienced it during the "lobster wave". In fact, besides WorkBuddy, Tencent also has CodeBuddy for developers, which basically corresponds to Claude Cowork and Code.

Tencent has already used it within the company for R & D and non - technical teams. The official disclosure shows that CodeBuddy covers more than 95% of Tencent's engineers, and WorkBuddy is used for mixed - programming development of humans and AI and helps small teams iterate products more quickly.

Image source: Tencent

We won't go into details about CodeBuddy here. Compared with code development, WorkBuddy is oriented towards a wider range of productivity scenarios, responsible for document processing, content organization, collaborative tasks, and general office needs.

This is also Tencent's advantage. WorkBuddy can fully utilize the ecological advantages of Tencent Docs, Tencent Meeting, ima Knowledge Base, Tencent LeXiang, and WeChat.

Many individual users' materials, meeting minutes, to - do lists, and communications are already scattered in these platforms. So an important reason for many people to use WorkBuddy is that it can directly pick up the daily "context" and then organize content, generate documents, and advance tasks in WorkBuddy.

Kimi Code/Work: Incorporating long - text into the Agents workflow

Facing the wave of Agents, a group of large - model companies that recently launched corresponding products have also divided their offerings into Kimi Code and Work.

Kimi Code can enter the CLI and IDE, read and write files, execute commands, search for code, obtain web content, and generate sub - agents for parallel tasks; Kimi Work is oriented towards the local workflow, can mount local folders, browse web pages through WebBridge, run Python, execute scheduled tasks, and requires user confirmation before modifying files or running code.

Image source: Kimi

For a long time in the past, Kimi has left the impression of "long - text processing" on ordinary users. Reading papers, financial reports, dozens of pages of PDFs, and organizing a large amount of web materials are the capabilities of Kimi that are most easily perceived by users. In the Agent stage, what Kimi needs to do is also very natural: not only read the materials and answer questions, but also continue to help users process files, run scripts, modify code, and generate results.

If the strength of Claude Code and Codex lies in the closed - loop of code tasks, what is more worthy of attention about Kimi is that it is trying to turn the advantage of "long - text reading" into the advantage of "long - range tasks". At the same time, in both Kimi Code and Work, the design of an Agent cluster is also inevitable. According to the official statement, when dealing with complex problems, Kimi can automatically coordinate multiple professional agents and simultaneously decompose and solve multi - level tasks.

In addition, to attract financial users, Kimi Work also pre - integrates in - depth data sources of A - shares, Hong Kong stocks, and US stocks. All these can be said to constitute the differentiated experience of Kimi Code and Work.

Alibaba Qoder: Growing from Tongyi Lingma into a desktop Agent

Alibaba Qoder is a bit special. It is not a brand - new product that emerged from scratch, but evolved from an AI IDE product (Tongyi Lingma) through several rounds of iteration into a new Agent product, and has derived a series of products such as Qoder Desktop, QoderWork, QoderWake, Qoder CLI, and Cloud Agents.

Image source: Alibaba Qoder

The core lies in two products: Qoder Desktop and QoderWork. Qoder Desktop is oriented towards software development scenarios and is closer to Claude Code; QoderWork is oriented towards daily work, handling tasks such as file organization, data processing, document generation, browser automation, desktop control, and scheduled tasks, and its positioning is similar to Claude Cowork.

QoderWork transfers Agent capabilities from code to ordinary work. As a desktop intelligent work assistant, QoderWork can complete file organization, data processing, and document generation through natural language, and can also connect to office tools, control browsers and computers; it also supports scheduled tasks, which is suitable for repetitive but indispensable work such as daily data pulling, weekly report writing, and monthly material organization.

From the product design perspective, in addition to the "general" mode, QoderWork also provides "design", "slide", and "writing" modes, which shows obvious emphasis on actual user scenarios. However, apart from this, not many features and advantages can be seen.

MiniMax Code/Agent: Vertical integration of model and Agent

Like the previous companies, MiniMax has launched two products: MiniMax Code and MiniMax Agent. It is worth mentioning that with the release of the new - generation large model MiniMax M3 in June, MiniMax Code also underwent a major update. According to the official statement, it is an Agent product specifically designed for M3 and trained together with M3.

Image source: MiniMax

In short, MiniMax Code can fully leverage M3's capabilities in long - context, Coding/Agentic, and native multi - modality, and is the preferred Agent to be paired with MiniMax - M3. In terms of long - range tasks, MiniMax Code will break down tasks into Workflows, which are coordinated by an Agent cluster and autonomously run through the adversarial Harness cycle of Producer + Verifier.

In fact, Claude Code has also launched Dynamic Workflows with a similar strategy, but MiniMax Code focuses more on "deep reflection and continuous error correction". The Agent will adjust the plan and priority in real - time according to the task progress, and users can also intervene at any time to add requirements or correct the direction.

As for MiniMax Agent, it actually tried to tackle long - range tasks earlier. In the May update, it launched the Mavis mode, which uses the collaboration of multiple Agents, including a design similar to Codex that allows users to intervene in the Agent's thinking and work at any time. It can be regarded as a preview before the release of MiniMax M3 and the update of MiniMax Code in June.

However, it should also be noted that this path is most likely to have a gap between the demonstration and the real experience. Long - context does not mean truly understanding the whole picture, and multi - Agent collaboration does not necessarily mean more stable results. The more roles and the longer the chain, any deviation in the middle may be magnified into an error.

ByteDance TRAE Work: Moving from an AI IDE to a general workbench

Strictly speaking, TRAE (SOLO) was first recognized by more people because it is an AI IDE, competing with development tools such as Cursor and Claude Code. However, the recently upgraded TRAE Work has expanded its boundaries, and its official positioning is very clear: not just for coding.

According to the official introduction, TRAE Work provides multi - terminal entrances on the Web, Desktop, and Mobile, takes into account both local and cloud tasks, and does not depend on the TRAE IDE to run. After users initiate tasks on the desktop, the Agent can continue to run on the cloud or locally, and multiple tasks can be advanced in parallel; after leaving the computer, users can also check the progress, review the results, and continue to adjust the direction on their mobile phones.

Image source: ByteDance TRAE

It is also worth mentioning that TRAE Work is divided into two modes: Work and Code. You can simply understand it as ByteDance's Codex, which combines the two Agent products, Code and Work, into one. The Code mode continues to handle development tasks, while the Work mode is oriented towards more general work scenarios, such as data organization, project promotion, web operation, file processing, and content generation. The difference is that Codex truly combines the two, while TRAE Work runs in two modes and requires users to enter from different mode entrances.

This is very much in line with ByteDance's style. Compared with Tencent WorkBuddy, which relies on the ecological context of WeChat, documents, and meetings, TRAE Work is more like an efficiency entrance for individuals and small teams. It does not necessarily require users to enter an office ecosystem first, but instead takes "tasks" as the core unit: as long as you give it your requirements, it is responsible for decomposition, execution, and feedback of progress, and users can intervene at any time during the process.

In addition, an important move of TRAE is the open - sourcing of Trae Agent. According to the GitHub page, Trae Agent is an LLM Agent toolkit for software engineering tasks, supporting file editing, bash execution, structured thinking, and task completion, and can also connect to multiple model providers.

Zhipu ZCode: Deeply exploring the Code Agent line first

Compared with other companies, Zhipu's approach is more focused. It is not in a hurry to package itself as a general - purpose office Agent, but clearly focuses on the developer scenario and launches Zhipu ZCode. Like Kimi and MiniMax, it aims at the difficulties and scenarios of long - range tasks. This is actually different from the thinking of many domestic Agent products. Tencent, Alibaba, Kimi, and ByteDance are all working on both the Code and Work lines.

It is worth mentioning that when Zhipu released the amazing GLM - 5.2 last month, it also upgraded ZCode 3.0 simultaneously, completely switching to the self - developed ZCode Agent kernel and deeply adapting to GLM - 5.2, and optimizing the long - range reasoning, tool call, and large - scale project execution link.

Image source: Zhipu

ZCode 3.0 also makes some improvements around the actual development experience, such as the grouped task workspace, the Zread project knowledge base, and the visual Git branch map. These may not sound as attractive as multi - Agent collaboration and long - context, but they are all related to the stability of the Code Agent: for an Agent to take over an engineering task, it first needs to understand the project, remember the context, and also let users clearly see what it has changed and where it is in the process.

The advantage of this approach is a clear goal, but the shortcoming is also obvious. The user base of ZCode will not be as large as that of WorkBuddy or TRAE Work for the time being, and its imagination space is more concentrated in the developer scenario. However, if it can truly run stably in complex code libraries, long - range tasks, and project verification, it will be easier to build the trust of professional users.

The opportunities for domestic Agents lie in sufficient differentiation

Looking at these products together, domestic Agents are not replacing Claude Code, Claude Cowork, and Codex in the same way.

Alibaba Qoder, ByteDance TRAE, and Tencent's Buddy series from large companies all learn from Anthropic and OpenAI in the Code + Work model, but their product paths and focuses are significantly different. As large - model startup companies that have survived the "hundred - model war", Kimi, MiniMax, and Zhipu all pay more attention to technological advantages, try to overcome the "long - range task" that Agents must face, and emphasize the vertical integration from the model to the Agent.

But in the end, what domestic Agents are chasing is not just a single product like Claude Code, Claude Cowork, or Codex, but a brand - new work mode: users no longer just ask AI a single question, but hand over a task to AI and then review, correct, take over, and continue to give commands during the process.

Compared with overseas products, the first advantage of domestic Agents is that they are closer to the local workflow.

Claude Code, Claude Cowork, and Codex have a higher