Abstract
The rise of AI agents is reshaping both software development and daily office work. After Andrej Karpathy popularized the concept of Vibe Coding, natural-language-driven workflows began moving beyond programming. Kimi Work is a new example of this shift. It extends the same interaction model into mainstream office scenarios and introduces a broader concept: Vibe Working.
Kimi Work is built around three key capabilities: WebBridge, large-scale agent clusters and a flexible skill ecosystem. Together, these modules allow the platform to automate document processing, web operations and collaborative office tasks.
One of the most notable facts is its development process. More than 92% of Kimi Work’s code was generated by AI, and the product was developed and launched in just one week. This makes it a strong example of the “AI building AI” trend.
This article analyzes Kimi Work’s product positioning, technical architecture, core functions, development process, security design, practical scenarios and industry impact. It also reviews how this product reflects a wider transition from AI-assisted coding to AI-assisted working.
1. Industry Background and Product Positioning
In recent years, Vibe Coding has become a recognizable development pattern in the technology industry. Developers can describe goals in natural language, and AI agents then complete coding tasks, modify files and run project-level operations. This has lowered the barrier to software creation.
However, programming is only one part of modern work. Most office users spend far more time on documents, spreadsheets, communication and business processes. According to macOS application usage data, 67.2% of daily active device usage is related to document editing, spreadsheet processing and business communication, not programming. This creates a much larger market for AI tools designed for general office work.
Kimi Work was launched against this background. Its goal is to extend the logic of Vibe Coding into broader office scenarios. Instead of writing code, users describe work objectives in natural language. The platform then executes a series of linked tasks automatically.
This makes Kimi Work different from typical AI coding products. Its main focus is not software development, but daily office productivity. It is designed for document handling, web interaction, information processing and collaborative work.
The product also comes with professional industry data sources. This is especially useful for financial users, who often need quick access to structured business and market information. In these scenarios, Kimi Work aims to provide an out-of-the-box experience with less configuration.
At the same time, this new work paradigm brings new questions. AI automation can make work more efficient, but it may also make work more continuous. When tasks become easier to start and execute, some users may end up working longer hours. This means enterprises need not only new tools, but also new usage norms.
Kimi Work is therefore more than an office product. It is an early attempt to explore what happens when AI agents become collaborative team members on ordinary office devices.
2. Core Technical Architecture: WebBridge and Agent Clusters
Kimi Work relies on two core technical modules: WebBridge and large-scale agent clusters. These two modules can be understood as the platform’s “working hands.” They allow the system to interact with web pages and handle complex tasks in parallel.
This architecture moves beyond the limitations of traditional single-agent systems. It improves both the scope and efficiency of task execution.
2.1 WebBridge: Human-Like Browser Operations
Kimi first introduced WebBridge in mid-May 2026 and later integrated it into Kimi Work.
Traditional AI agents often struggle with web operations. Many tools can only call fixed APIs. They cannot interact with websites in the same way a human user does. This limits their usefulness in real office and operational scenarios.
WebBridge addresses this problem. It allows Kimi Work to operate mainstream browsers more like a real user. The system can perform common web interactions, such as clicking, navigating, filling in information and managing page-based tasks.
In practical scenarios, users can give simple instructions and ask Kimi Work to complete actions on social platforms such as X. For example, it can support tasks like liking, following, unfollowing or interacting with posts.
This extends AI agents beyond text generation. It turns them into tools that can perform interactive web operations. For marketing and operations teams, this opens up use cases such as social media maintenance, account management and routine web-based workflows.
2.2 Large-Scale Agent Clusters for Parallel Execution
Kimi Work also inherits agent cluster technology from the Kimi K2.5 model. This allows the system to split a complex task into smaller subtasks and assign them to multiple agents.
The platform can mobilize up to 300 agents at the same time. This is a major improvement compared with traditional single-agent tools.
The advantage is clear in large tasks. A single agent must usually process work step by step. An agent cluster can handle different modules at the same time. This reduces waiting time and shortens the overall task cycle.
A practical test used the prospectus PDF of Changxin as source material. The task was to generate a PPT based on the document. The agent cluster completed the task much faster than a single AI agent.
The key reason is parallel execution. Different agents can process different pages, sections or content modules simultaneously. After that, the system aggregates the results into a final output.
This architecture is especially suitable for large documents, multi-page presentations and complex office tasks. It also supports hierarchical scheduling and result aggregation, which helps maintain output consistency.
3. Skill Ecosystem: Flexible Functional Expansion
Kimi Work includes a Skill Square, which functions as the platform’s extension ecosystem. It allows users to activate, install and call different skills based on their needs.
The Skill Square is divided into two main sections: installed skills and recommended skills. This design helps users distinguish between available functions and newly suggested capabilities.
Users can activate pre-installed skills directly. These skills cover common office tasks, knowledge search and scenario-based services. If the built-in functions are not enough, users can also install custom local skills.
There are two ways to trigger a skill. Users can select it directly from the Skill Square, or they can call it through prompts in the dialogue box. This gives users flexibility in different workflows.
The pre-configured Zanggui PPT skill is a typical example. It allows users to quickly generate standard presentation content without complex setup. This reflects Kimi Work’s focus on practical office productivity.
However, the Skill Square still has room for improvement. Public feedback has mentioned the need for better data statistics, category management, content search and popularity rankings. More Chinese labels for installed skills would also improve the experience for domestic users.
As the skill ecosystem grows, Kimi Work may cover more specialized office scenarios. The long-term value of the platform will depend not only on its base model, but also on the richness of its skill ecosystem.
4. Development Process: AI Building AI with 92% AI-Generated Code
The development process of Kimi Work is one of its most important stories.
The product was completed and launched in only one week. During development, Kimi engineers used tools such as Kimi Code to write and generate program code. The total effective codebase exceeded 50,000 lines. More than 92% of the code was generated by AI.
This makes Kimi Work a representative case of “AI building AI.” It shows that AI is no longer only used to assist individual developers. It can now participate deeply in the creation of full software products.
The underlying model behind Kimi Work is Kimi K2.6. Compared with earlier versions, K2.6 is optimized for long-cycle tasks and agent cluster scheduling. These capabilities are important for office automation, where tasks often involve multiple steps, documents and external systems.
Many third-party agent products face stability issues when handling complex work. They may fail midway, lose context or interrupt long tasks. Kimi Work has an advantage as a first-party product built on its own model stack. The model and client are designed to work closely together.
This improves end-to-end stability. It also helps the product deliver a smoother execution experience in long and complex tasks.
5. Security Mechanism and Permission Management
Security is critical for any office AI tool. Kimi Work may process internal documents, business data and sensitive files. If access control is weak, the risk of data leakage increases.
To reduce this risk, Kimi Work uses file isolation and permission control. By default, tasks run only within designated folders. This prevents the AI agent from freely accessing the entire local disk.
The platform also provides two permission modes.
The first mode requires the agent to request user permission before performing high-risk operations. This is safer for sensitive work.
The second mode gives the agent full access to a designated working directory. This improves efficiency when users trust the task environment.
This two-level permission design balances safety and convenience. Individual users can choose stricter controls for private files. Enterprise teams can define permission rules based on document sensitivity and internal security policies.
For AI agents deployed on ordinary office devices, this type of permission management is essential. It helps reduce data leakage, accidental file changes and unauthorized access.
6. Application Scenarios and User Value
Kimi Work covers a wide range of office scenarios. Its users can include individual employees, financial professionals, marketing teams and enterprise departments.
For ordinary office workers, Kimi Work can automate document processing, spreadsheet handling and daily task management. Users only need to describe the goal. The platform can then execute related steps and reduce repetitive manual work.
For financial practitioners, built-in professional data sources are especially useful. Users can quickly collect industry information, summarize reports and analyze structured business data.
For operations and marketing teams, WebBridge creates more possibilities. It supports web-based operations such as social media interaction, account maintenance and content management. This allows teams to automate some routine online workflows.
For team collaboration, agent clusters can split large tasks into smaller parts and process them in parallel. This can improve efficiency in projects that involve multiple documents, departments or output formats.
Compared with single-function AI tools, Kimi Work combines web operation, document processing and intelligent scheduling. This makes it closer to a one-stop office automation platform.
For teams using multiple AI models and agent tools, an API gateway such as 4sapi can be evaluated as a supplementary access layer. It may help simplify multi-model invocation and reduce integration complexity.
7. Existing Limitations and Industry Trends
Kimi Work has strong functional potential, but the current version still has limitations.
The Skill Square is one example. It still lacks mature category management, search functions and popularity statistics. This can make it harder for users to find the right skill quickly.
The localization experience can also improve. More Chinese labels and clearer descriptions for installed skills would help domestic users understand available functions faster.
These issues do not weaken the product’s overall direction. But they do show where later versions need to improve.
From an industry perspective, Kimi Work marks the arrival of the Vibe Working concept. Office work is starting to follow the same natural-language-driven path as AI-assisted coding.
In the future, more office tasks will be completed by AI agents. Users will give goals, not step-by-step instructions. The boundary between manual work and intelligent automation will continue to blur.
This will also intensify competition. Future AI office products will not compete only on basic functions. They will compete on scenario adaptation, security design, ecosystem depth and task stability.
There is also a social dimension. More efficient tools may make work easier, but they may also extend work time. Enterprises and individuals need clear rules for using AI tools. Efficiency gains should not come at the cost of unhealthy working patterns.
8. Conclusion
Kimi Work is a landmark product in the shift from Vibe Coding to Vibe Working.
It was launched in just one week, with 92% of its code generated by AI. This makes it a strong example of the “AI building AI” development model.
Its core capabilities include WebBridge, large-scale agent clusters and a flexible skill ecosystem. WebBridge allows the platform to perform browser-based operations. Agent clusters support parallel processing of complex tasks. The skill system gives users a way to expand functions for different office scenarios.
Kimi Work also shows the importance of security in AI office products. Its folder isolation and permission controls help reduce risks when agents handle local files and business data.
The product is not perfect. The Skill Square still needs better search, classification and localization. But its architecture and product logic point to a clear direction for the AI office market.
In the future, natural-language-driven work will become more common. AI agents will move from assisting individual tasks to supporting entire workflows. Products like Kimi Work may reshape how people create documents, manage information and collaborate across teams.
The next stage of AI office competition will not be about one single feature. It will be a broader contest of ecosystem strength, security, scenario depth and long-task reliability.
