Kimi Work Review: 300 Parallel Agents Power Automated Stock Analysis

Kimi Work Overview

Kimi Work is the desktop agent mode of the Kimi computer client, distinct from the chat‑oriented web version. It can read local files, open browsers to scrape pages, write documents, schedule tasks, access professional data sources, and split complex jobs into up to 300 parallel sub‑agents.

Installation is available for macOS (Apple silicon) and Windows at https://www.kimi.com/zh-cn/products/kimi-work . After selecting Work mode and enabling the K2.6 Agent cluster, the user can start processing desktop tasks.

300 Parallel Agents in Action

AI Hotspot Radar

The author creates a timed radar that scans AI‑related sources (GitHub Trending, Hugging Face, Product Hunt, Hacker News, major AI company blogs) every 24 hours, filters results on relevance, playability, writing value, and risk, and outputs a markdown report.

请执行 AI 开源项目每日扫描任务：
1. 并行扫描以下平台过去 24 小时内的 AI 开源项目动态：
   - GitHub Trending（github.com/trending，关注 AI/ML 相关仓库）
   - Hugging Face（huggingface.co/spaces 和热门模型）
   - Product Hunt（producthunt.com，筛选 AI 开发者工具）
   - Hacker News（news.ycombinator.com，搜索 AI/ML 相关热门帖）
   - 主流 AI 公司博客（OpenAI/Anthropic/Google DeepMind/Meta AI/Mistral/Cohere 的最新发布）
2. 按以下四维度筛选 Top 5（每项 1-5 分）：
   - 程序员相关性：是否直接面向开发者，是否有代码/工具属性
   - 可玩性：能否在 1-2 小时内跑通最小 Demo
   - 写作价值：是否有技术深度、独特视角、可产出文章
   - 避坑指数：文档是否完善、社区是否活跃、依赖是否复杂（分数越高越不容易踩坑）
3. 为 Top 1 项目生成 1-2 小时实测任务单，包括：
   - 安装步骤（必须可验证，不确定写"待核实"）
   - 最小 Demo 命令/代码
   - 测试任务清单
   - 截图清单
   - 可能踩坑点
   - 建议文章标题
   - 待核实信息列表
4. 输出要求：
   - 所有关键结论必须给来源链接
   - 不确定的地方写"待核实"
   - 不要编造命令、参数或性能数据
   - 不做 AI 新闻日报，只聚焦可动手体验的开源项目
   - 最终输出保存为 Markdown 文件到当前工作区

请使用多个子代理并行完成扫描，然后整合结果。

Kimi Work launches multiple SubAgents to process the radar, as shown in the execution flow image.

The final markdown report contains the filtered AI projects.

Kimi Work also supports scheduled tasks; the author sets the radar to run daily at 8:30 AM so results are ready by 9:00 AM.

Stock Analysis Demo

Using built‑in professional data sources (Tonghuashun, ifind, Tianyancha, Yahoo Finance, World Bank, Binance, arXiv, Google Scholar), the author simulates a six‑person investment research team with six parallel agents: Researcher, Fundamental Analyst, Technical Analyst, Sentiment Analyst, Risk Officer, and Portfolio Manager.

请你模拟一个股票投研团队，对【股票名称/代码】进行一次"AI 投资委员会"分析。

注意：
本任务仅用于研究和产品能力展示，不构成任何真实投资建议。
请不要直接输出"买入/卖出/加仓/清仓"等交易指令。
所有结论必须有数据来源，无法确认的信息标记为"待核实"。

请启动多个子 Agent 并行工作，角色如下：

## 角色一：研究员 Researcher
职责：收集该股票最近的行情、公告、财报、研报摘要、行业新闻和公司动态。
只做客观信息聚合，不做主观投资判断。

## 角色二：基本面分析师 Fundamental Analyst
职责：分析公司的收入结构、利润变化、毛利率、现金流、估值、行业竞争格局。
判断基本面是否出现变化，但不要给交易建议。

## 角色三：技术分析师 Technical Analyst
职责：分析股价走势、K 线、均线、MACD、RSI、KDJ、布林带、成交量。
识别关键支撑位、压力位和趋势状态。

## 角色四：舆情分析师 Sentiment Analyst
职责：分析财经新闻评论、社交平台讨论中的市场情绪。
识别散户情绪、市场热度、争议焦点和潜在误读。

## 角色五：风险官 Risk Officer
职责：专门寻找反面证据和风险事件。
包括监管风险、诉讼风险、股东减持、限售解禁、业绩不及预期等。

## 角色六：投资经理 Portfolio Manager
职责：不直接查询新数据，只阅读以上五位角色的报告，组织一次投资委员会总结。

最终交付物：
1. 一份 Word 总报告；
2. 一张 Excel 数据来源与证据表；
3. 一份 Markdown 投委会纪要；
4. 一份风险清单；
5. 一页纸投资经理总结。

The execution screenshots show each role producing its report, followed by the Portfolio Manager’s consolidated minutes.

Custom Stock‑Analysis Skill

The author packages the prompt into a reusable Skill, allowing one‑click execution for different stocks. Skills can be built‑in, installed from a marketplace, or custom‑defined.

Skill generation steps are illustrated in the diagram below.

WebBridge Browser Automation

Kimi Work’s WebBridge extension enables direct browser control. After installing the extension, the agent can navigate to a public account’s backend, collect article statistics, and output aggregated results.

Pros and Cons

WebBridge browser control greatly aids information‑intensive tasks.

300 parallel agents are effective; splitting work among dedicated agents improves stability.

Professional data sources provide higher accuracy than pure web scraping.

Skill persistence lets users reuse prompts without rewriting.

Speed is slower than Codex or Claude Code, and long tasks can be unstable.

Lack of a planning mode means tasks start immediately without prior scope confirmation.

Occasional instability in long browser sessions may require manual intervention.

Overall, Kimi Work automates many repetitive knowledge‑work steps but still requires user oversight for factual correctness and task stability.

Conclusion

The product illustrates a shift from asking AI “how to do this?” to instructing AI “here are the materials, process them and deliver a complete result.” While still in beta, its combination of massive parallel agents, WebBridge, local file access, scheduled tasks, professional data sources, and Skill encapsulation makes it a compelling prototype for AI‑augmented work.