Kimi Work to Run 300 AI Agents in Parallel

OpenAI’s Codex was once a code completion tool for developers, but now has over 5 million weekly active users, with knowledge workers making up about 20% of its user base. Following this trend, Chinese AI startup Moon’s Dark Side (Moonshot AI) has launched Kimi Work. Kimi Work is an agent cluster that operates on local PCs, capable of running up to 300 AI agents in parallel.

According to an OpenAI report, developers still form the largest user group of Codex, but the growth rate of knowledge workers is more than three times that of developers. Claude Code offers similar functionality, but its installation and deployment barriers have limited its adoption among general users. Kimi Work breaks down this wall with the ease of one-click installation.

The essence of Kimi Work is the evolution from a single agent (AI assistant) to an “Agent Team,” and further to a dynamic workflow consisting of up to hundreds of agents. Moon’s Dark Side achieved this complex mechanism by encapsulating the capabilities of the underlying Agent Swarm model in a visual interface and deploying it on local PCs.

Overview of the Local AI Legion

Kimi Work’s greatest advantage lies in configuring Agent cluster functionality as a local agent. To make Agent Swarm truly functional, Kimi Work provides several practical features.

• Deep integration with local files: Can directly read and manage local folders. A security mechanism requires user approval for file changes.
• 7×24 scheduled tasks (Cron engine): Automatically executes LLM interaction requests, Python/Shell scripts, etc., at set times. Ensures tasks run reliably, such as generating morning briefings or cleaning data at night, preventing the PC from sleeping.
• WebBridge browser automation: AI autonomously controls the browser via natural language instructions. Performs web searches, deep data scraping, automatic form filling, etc.
• Native access to global financial market data: Directly connects to major data sources such as A-shares, Hong Kong stocks, and US stocks. Can fetch financial statements, chart analysis, cross-sheet comparisons during conversations, aiding investment decisions.

These features significantly expand the domain of traditional AI agents, which were limited to “writing code.” Specifically, they can handle file organization, data analysis, document retrieval, automated workflow execution, and even entire project iteration tasks.

Examples of Agent Swarm in Action

To see Kimi Work in action, there is an example where it was asked to organize 20 notable AI companies, analyze their product positioning, funding trends, core competencies, etc., and output deliverables such as web reports and PPTs.

After selecting the K2.6 Agent cluster and starting the task, Kimi automatically sets an appropriate progress for the task and uses Subagent tools to call multiple agents for processing. Expanding the task process reveals four research agents: Research Group 1 (major AI companies), Research Group 2 (emerging large models), Research Group 3 (infrastructure), and Research Group 4 (AI applications), each collecting and analyzing materials on their respective companies.

Skill invocation is a basic operation; Kimi calls skills like report generation, visualization, cluster deep research, and frontend themes to assist the task. The final analysis report included data tables, visualization analyses, and specific company introductions. Additionally, each company’s details included a “Risk” section, where Kimi’s self-assessment noted “valuation rising too fast, unclear monetization inflection point” and core competencies as “long-text processing, programming ability, agent, and open-source leadership,” reflecting its own evaluation.

Beyond deep research tasks, Kimi Work can also directly process local files. For example, it can be asked to organize files from the last 30 days, use tools to check file contents, and aggregate file information. From the progress on the right, it reads text file contents, extracts PDF files, checks image files, processes Office files, and finally generates a table listing all file contents and corresponding information successfully.

If you provide a meeting memo, some thesis materials, and image data, resulting in over 10 local files, Kimi Work can directly select that folder as a project and start Agent Swarm processing. Using the K2.6 Agent cluster for parallel collaboration, you can ask it to create multiple documents based on the folder materials, such as industry research, thesis reviews, product strategies, technical architecture, compliance governance, financial projections, PPT design, Word reports, PDF research reports, Excel models, and quality reviews.

Due to the large number of files, the Kimi K2.6 Agent cluster executes in multiple phases. In the first phase of research analysis, it calls six agents: industry researcher, thesis reviewer, product strategist, technical architect, compliance governance expert, and financial projection specialist. In the second phase, it calls four agents: PPT designer, Word report creator, PDF researcher, and Excel modeler, to integrate and deliver. After integration and delivery, quality review automatically begins, with two subagents—quality reviewer and web developer—final-checking the previous content.

Ultimately, Kimi Work generated six text reports, one projection model, and HTML and PPT documents for reporting. For each file, quick navigation is provided based on usage scenarios: “07_Management Report.pptx” + “12_Digital Report.html” for reporting to CEOs/management, “08_Comprehensive Consulting Report.docx” for formal consulting report submissions, “09_In-Depth Research Report.pdf” for investment banks/research institutions, etc.

Changes in Knowledge Work Patterns

Traditionally, knowledge workers followed a “one person processes tasks sequentially” model. However, with the advent of local agents like Kimi Work, knowledge work is shifting toward “one person commands a group of AIs to process tasks.”

Beyond directly using the capabilities of 300 agents, combining them with Agent clusters leverages Kimi’s built-in unique financial data sources. There is no need to manually search for financial-related skills or set up data APIs; Kimi directly fetches financial data from sources like Tonghuashun, Tianyancha, and the World Bank economic database.

Combining this data with Agent clusters unleashes true power. Just before Apple’s WWDC, when asked to compile Apple’s stock price information over the past three years and annual financial statements, and analyze noteworthy information, Kimi similarly launched Subagent tools, calling multiple agents to complete the task.

The emergence of Kimi Work demonstrates that AI agents are evolving from simple code completion tools into universal productivity platforms. Specifically, bringing the concept of Agent Swarm into the local environment has the potential to significantly change how knowledge work is done in enterprises. As mentioned in this site’s previous article “What is an AI Agent? Explanation of Mechanism and Major Frameworks,” the practical application of inter-agent collaboration and parallel processing is making the automation of large-scale tasks—difficult with single agents—a reality.

Additionally, cost optimization is a key consideration when introducing AI agents. As discussed in this site’s article “AI Agent Cost Optimization: Practical Techniques to Reduce Token Consumption,” running many agents in parallel can lead to a surge in token usage. How Kimi Work addresses this challenge remains to be verified.

Editorial Opinion

Short-term Impact: The emergence of local Agent Swarms like Kimi Work is expected to directly impact enterprise business processes within the next 3 to 6 months. In particular, automation of routine tasks such as file organization, scheduled tasks, and simple research will likely proceed, contributing to reducing the burden on knowledge workers. At the same time, competitors like Microsoft and Google will likely accelerate moves to strengthen similar local agent functions. However, at present, many features are tailored for the Chinese market (A-share, Hong Kong stock data, etc.), and localization will be a challenge for global expansion.

Long-term Perspective: Over a 1–3 year span, the parallel processing capability of Agent Swarms could change the nature of knowledge work. Business processes traditionally designed on the premise of “humans processing sequentially” will be redesigned into a model where “humans command AI teams.” This is expected to dramatically expand the scope of work a single knowledge worker can cover, flattening organizations and increasing flexibility in talent allocation. On the other hand, new skills (AI orchestration) will be required to properly manage and supervise hundreds of agents, posing a risk of creating a new digital divide.

Editorial Inquiry: Are the costs and management complexity of running numerous AI agents in parallel acceptable for small to medium-sized enterprises and individual users? The true value of Agent Swarm depends on the quality of the experience where “humans give a single command, and multiple AIs automatically collaborate to produce results.” In particular, the governance mechanism for ensuring the consistency and quality of information generated by each agent remains unclear. Readers are encouraged to experiment with how to incorporate this “distributed AI team” into their own work.

References

• iFanr article “300 AIs and My Six-Pack Abs” — published June 8, 2026
• OpenAI Codex related information (official reports, etc.)