文 | 一点财经，作者 | 赵同，编辑 | 邹珺 它究竟带走了哪些秘密？ 大闹一场，悄然而去 当产品内测邀请码被炒到10万元一个之后，Manus突然就"跑路"去了新加坡。 它也成为首个"跑"去海外的有名气的AI公司，那些满怀期待的用户、希望用炒作赚钱的黄牛，瞬间坠入 冰窖。 今年3月，由中国团队打造的Manus上线，一度被热捧为"下一个DeepSeek"——不是只会聊天的AI工 具，而是能直接做任务然后交出结果的智能体。 这点燃了许多人的热情，一夜之间Manus被形容为"国运级产品"，仿佛不用它就是不爱国。内测邀请码 炒作到10万元一个，比周杰伦的演唱会门票更贵、更难抢。 狂热有多猛烈，冷却就有多迅速。又是一夜之间，最近Manus国内的社交媒体清空了，网页无法登陆、 团队全部裁员，一些合作也结束了，创始人肖弘带着核心成员奔向了新加坡。 从众星捧月到远走他乡，Manus只用了4个月。没有公告，没有解释，甚至没有一句江湖再见。留下满 屏问号：昔日被捧上"国运"神坛的宠儿，为何转身如此决绝？是无奈出走，还是早有预谋？ 现在市面上有两种说法，一种是怒骂Manus之前疯狂割韭菜，割完就跑了，表示愤懑。另一种认为 Man ...

还记得三个月前那个让全网欢呼的AI Agent吗？ 2025年3月，Manus横空出世，凭借一段"智能体自主完成任务"的演示视频，一夜之间成为科技圈宠儿。发布一周内，200万用户挤破头预约，内测码被炒 至10万，仿佛AI的下一个ChatGPT时刻已经到来。 实际上，Manus的困境并非孤例。另一家明星Agent企业澜码科技早在2025年初就因融资断裂，停发员工薪资数月，目前正寻求并购机会。 虽然通用智能体被广泛认为是实现AGI的必经之路，但现实却给了这个美好愿景沉重一击。行业报告预测，到2027年，约40%的AI Agent项目可能因成本 失控或商业模式不清晰而被淘汰。 那么，通用Agent赛道的窘境，是否折射了整个Agent行业的困局？未来的Agent要怎么做，才能被市场买单？ 01 通用智能体：从被追捧到被群嘲 在让Manus一夜爆红的那段视频里，智能体能自动筛选简历、分析股票、规划旅行，甚至能像人类一样"思考"复杂任务。此后，内测邀请码一码难求，资 本蜂拥而至，硅谷顶级风投Benchmark领投7500万美元，公司估值飙升至5亿美元。媒体争相报道，称其为下一代人机协作的范式。 一时间，Manus成了AI ...

这个来之不易的教训让我们的选择变得异常清晰：Manus 将赌注押在 上下文工程（Context Engineering） 上。这使我们能将产品改进的周期从数 周缩短至几小时，并让我们的产品与底层模型的发展保持正交：如果说模型技术的进步是上涨的潮水，我们希望 Manus 是水涨船高的船，而非被固定 在海床上的桥墩。 即便如此，上下文工程的实践之路也远非一帆风顺。它是一门实验科学——我们已经重构了四次 Agent 框架，每一次都是因为发现了塑造上下文的更 优方法。我们将这种手动进行架构搜索、调试提示（prompt）和凭经验猜测的摸索过程，戏称为" 随机研究生下降 "（Stochastic Graduate Descent）——这是对"随机梯度下降"（Stochastic Gradient Descent）的一个文字游戏。这个方法听起来不那么优雅，但确实有效。 本文是对 Manus 创始人 季逸超 （Jiyichao）在其博客上发表的《Context Engineering for AI Agents: Lessons from Building Manus》一文的全文翻译。 这篇文章并非空泛的理论，而是 Manu ...

作者 | 季逸超 Peak Ji，蝴蝶效应联合创始人、首席科学家 译者 | 王强 策划 | Tina 在 Manus 项目伊始，我和我的团队面临一个关键决策：我们是应该使用开源基础模型训练一个端到 端的 Agent 模型，还是应该在前沿模型的上下文学习能力之上构建一个 Agent？ 之前我还在 NLP 领域的第一个十年时，我们是没有这样的选择余裕的。在 BERT（是的，已经过去 七年了）那段遥远的日子里，模型必须先进行微调并评估，然后才能转移到新任务上。这个过程通常 需要每周迭代一次，尽管与今天的大型语言模型相比，当时的模型很小。对于快速进化的应用，尤其 是处于产品市场契合阶段（PMF）之前的应用，这样的 慢反馈循环 是不可接受的。这是我上一次创 业的苦涩教训，当时我从头开始训练模型，用于开放信息提取和语义搜索用例。然后 GPT-3 和 Flan- T5 出现了，我的内部模型一夜之间就落伍了。讽刺的是，这些模型标志着上下文学习时代的开始 ——而且开启了一条全新的前进道路。 那个来之不易的教训让我们下定决心： Manus 将押注于上下文工程 。这使我们能够在几小时内而不 是几周内发布改进，并且使我们的产品与底层 ...

Pig 当时正在攻克计算机使用这一关键领域，这是让 AI 代理真正能在职场发挥作用必须解决的重要 课题。另一家同样来自 YC 校友的公司 Browser Use ，正致力于解决浏览器端的同类问题。 当智能工具 Manus （依赖该技术） 爆红时， Browser Use 随之声名鹊起。 该技术本质上通过扫描 网站的按钮和元素，将其转化为更易消化的 " 类文本 " 格式，帮助 AI 理解如何浏览和使用网站。 在发布的 Y Combinator 播客中，合伙人 Tom Blomfield 将 Pig 比作 "Windows 桌面版的 Browser Use" 。这期节目还邀请了热门氛围编程初创公司 Replit 的创始人兼 CEO Amjad Masad 参与讨论。 Masad 、 Blomfield 和 YC 合伙人 David Lieb 当时正在讨论，计算机长时间使用（以小时而非分钟 计）仍然是智能代理面临的主要障碍。随着推理上下文窗口的扩大，代理的准确性会波动，同时 LLM 成本也会增加。 " 我现在给创业者的建议是，要么采用浏览器应用，要么通过 Pig 实现 Windows 自动化，然后尝试将 其应用到企 ...

Core Viewpoint - Manus, once a rising star in the AI agent space, has faced challenges in the Chinese market and has now shifted its focus to overseas operations, particularly Singapore, while reflecting on its technical journey and lessons learned [1][3][8]. Group 1: Manus's Journey and Market Strategy - Manus was initially celebrated for its product, which was dubbed the "world's first general-purpose agent," capable of executing complex tasks like resume screening and stock analysis [3]. - Following its rapid rise, Manus faced significant challenges, leading to its withdrawal from the Chinese market and a strategic shift to Singapore, where it has relocated its headquarters and core team [3][8]. - The company has undergone layoffs, with approximately 80 non-core employees being let go, as part of its strategy to enhance operational efficiency [3][8]. Group 2: Technical Insights and Lessons Learned - Manus's co-founder, Ji Yichao, emphasized the importance of context design over merely competing on model capabilities, highlighting that the future of AI agents lies in effective context engineering [1][4]. - The team experienced multiple adjustments to their agent framework, ultimately achieving a local optimal solution after four iterations [1][4]. - Key operational insights shared include the significance of KV-Cache hit rates, suggesting methods to improve cache efficiency, such as maintaining stable prompt prefixes and avoiding dynamic tool modifications [4][5]. Group 3: Industry Context and Competition - The AI agent sector is witnessing significant advancements, with competitors like OpenAI and Kimi launching new products that enhance automation capabilities [11][12]. - OpenAI's new "ChatGPT Agent" integrates various functionalities, showcasing advanced performance metrics in multiple benchmarks [11]. - Kimi's K2 model, designed as a universal cognitive engine for the next generation of agents, has also demonstrated superior performance in key areas [12]. - The competitive landscape is intensifying, with major players and startups alike focusing on the commercialization of AI agents, raising questions about how pure agent startups like Manus can sustain themselves amid fierce competition [12].

Core Viewpoint - Manus, once a rising star in the AI agent space, has faced significant challenges in the Chinese market, leading to its withdrawal and a shift in focus to overseas markets, particularly Singapore [1][4][8]. Group 1: Manus's Journey and Market Response - Manus was initially celebrated as the "world's first general-purpose agent product," capable of executing complex tasks like resume screening and stock analysis, which led to a surge in demand and high secondary market prices for its beta access [3][4]. - Despite its initial success, Manus announced its exit from the Chinese market within six months, citing operational efficiency and a strategic focus on core business development [4][6]. - The company has relocated its headquarters to Singapore, with a significant portion of its team moving, while non-core staff were laid off [3][4][8]. Group 2: Technical Insights and Lessons Learned - Manus's co-founder, Ji Yichao, emphasized the importance of context design over merely competing on model capabilities, reflecting on the lessons learned from previous entrepreneurial experiences [1][4]. - The team opted for "context engineering" using open-source or commercial large models instead of developing their own foundational models, which allowed for quicker product iterations [1][4][5]. - Key operational insights shared include the significance of KV-Cache hit rates, maintaining stable input prompts, and avoiding dynamic modifications to tool lists to enhance model performance [5]. Group 3: Industry Context and Competitive Landscape - The AI agent sector is witnessing rapid advancements, with major players like OpenAI and Kimi launching new products that enhance automation capabilities [10][11]. - OpenAI's new "ChatGPT Agent" integrates multiple functionalities, showcasing the competitive edge of foundational models in driving agent capabilities [10]. - The market is becoming increasingly crowded, raising questions about how pure AI agent startups like Manus can differentiate themselves amidst growing competition from established companies [11].

在构建通用型 AI Agent 的道路上，目前业界主要形成了两条技术路线：端到端训练和上下文工程。 模型厂商通常拥有丰富的模型训练经验，以及强大的自有闭源基础模型，因此更倾向于通过端到端训练，来充分发挥自身的独特优势。而最近几家通用Agent创业公 司却都选择了后一条路径。 Manus团队最近发表了一篇题为《AI代理的上下文工程：构建Manus的经验教训》的文章，作为上下文工程路线的代表性实践者，他们分享了自己的技术选择和实 战经验。 Manus AI的技术负责人季逸超，在文中坦诚地分享了选择上下文工程而非端到端训练的原因：上一次创业时，他花费大量时间训练的自研模型在GPT-3发布后一夜 之间变得毫无价值。这个惨痛教训让他深刻认识到，在大模型快速迭代的时代，构建系统时应该"成为涨潮中的船，而非固定在海床上的支柱"。 这个来之不易的教训让选择变得清晰：Manus将押注于上下文工程。这使我们能够在几小时内而不是几周内发布改进，并使我们的产品与底层模型保持正交：如果 模型进步是涨潮，我们希望Manus成为船，而不是固定在海床上的支柱。 尽管如此，上下文工程远非简单直接。这是一门实验科学——我们已经四次重建代理框架， ...

Core Insights - The excitement surrounding the release of OpenAI's ChatGPT agent is primarily felt by competing companies rather than end users, indicating a competitive landscape in the agent market [5][6]. - Companies like Manus and Genspark are actively comparing their products with ChatGPT, suggesting a fierce competition and positioning themselves as superior alternatives [1][4][50]. Product Comparisons - Manus has released multiple tweets highlighting its agent's capabilities compared to OpenAI's, claiming to be faster and more efficient [1]. - Genspark showcased a demo that emphasizes its agent's ability to complete tasks more smoothly than ChatGPT, indicating a focus on user experience [4]. - The ChatGPT agent has been rolled out to Pro users, with demand exceeding expectations, leading to a phased rollout for Plus and Team users [6]. User Experience and Performance - A user tested the ChatGPT agent by generating a comprehensive retirement plan presentation, which took about 20 minutes to complete, but the final product was deemed simplistic [12][14]. - The agent's process involved automatic information gathering without user intervention, showcasing its efficiency [13]. - Comparisons with Manus and Genspark revealed that while ChatGPT can generate presentations, the quality and aesthetics of the outputs from competitors were often superior [50][105]. Market Dynamics - The launch of the ChatGPT agent is perceived as a significant event in the agent market, akin to a "competitive bomb" being dropped, which has prompted other companies to enhance their offerings [5]. - The competitive landscape is characterized by rapid responses from companies like Manus and Genspark, who are eager to demonstrate their products' advantages over ChatGPT [1][4][50]. Financial Independence and Retirement Planning - The article discusses a financial independence model (FIRE) for a high-income individual aiming to retire at 30 with $5 million, highlighting the challenges of achieving such goals in a high-cost city like Vancouver [156][160]. - The analysis indicates that even with high savings rates (80-90%), the target of $5 million may not be feasible without extraordinary investment returns or additional income sources [157][159].

在爆火仅四个月后，Manus AI 突然几乎全面撤出中国市场，不仅清空全部社交账号内容，而且国行版本的 Manus 也疑似暂停推进。 早在上个月，Manus 联合创始人张涛便曾宣布，公司已将全球总部迁至新加坡，并在东京和加州设有办公室。尽管官方未正面回应，只称是「基于经营效 率的调整」，但出海所引发裁员等一连串争议问题，也让外界普遍猜测其是否正在「跑路」。 风波之中，今天凌晨，Manus 联合创始人季逸超发布了一篇技术博客，试图将外界关注点重新拉回产品技术本身。 经过四次重构和数百万真实交互，他在文中坦诚地总结了团队在构建 Manus 过程中积累的经验教训。内容既有实操干货，也不乏反思，对业内同行与普 通用户来说，都不失为一份值得一读的参考材料。 1. 押注上下文，不再训练模型 与其耗时训练，不如围绕大模型构造「记忆」和流程。上下文工程让你在几小时而不是几周内发布产品更新。 2. KV-Cache 命中率至关重要 输入越稳定，缓存命中率越高，成本和延迟越低。三条实战建议： - 避免提示中使用时间戳； - 只追加上下文，避免修改历 史记录； - 手动标记缓存断点，保障前缀一致性。 3. 工具不要动态添加，而是用 ...