当下 AI 最大的机会，很多来自那些过去从未有软件能够解决的问题领域，或那些规模小到不值得开发软件的细分市场。 AI Agent 的出现，是第一次真正实现了“软件 + 工作”一起被交付给用户。这意味着即便是在传统上软件支出很小的行业，TAM（可服务市场规模）也会被放大得非常巨大——例如法律服务、医疗、会计、软件工程等等。 关键在于：这些市场往往需要对领域问题的深刻理解，因为 AI 不是单纯帮用户“启用一个工具”，而是真正参与到某条关键工作流中，成为增强劳动力的一部分。 像 Excel 这样的工具之所以能成为通用型技术，是因为用户自己能想办法用它完成各自独特的任务。 但在现实中，当你雇人来替你做一项工作时，你并不想亲自从头到尾训练他如何做所有的细节。当你今天雇佣劳动力时，你几乎不会希望对方是“通用型”的；你希望对方能解决你特定的问题。 AI Agents 也会如此。 用户会希望获得为特定领域或特定问题量身打造的 Agent，这将推动 Agent 按垂直行业或按特定工作流高度专业化。 因此真正的“游戏规则”是： 打造既具备深度领域知识，又能利用专有（客户）数据，并能紧密嵌入用户工作流（无论是现有的还是重新设计的），同时还能提供所有实施与变革支持资源，使客户真正成功的 AI Agents。 Many of the biggest opportunities in AI right now are going after categories where there was never software to accomplish the problem before, or the categories were too small to matter. AI agents represent the first time ever where we can bring along the work with the software, which means TAMs get very big even in spaces that traditionally had small software spend. Legal services, healthcare, accounting, software engineering, and so on. The key though is a lot of these markets require a deep understanding of the problem in the domain because you’re fundamentally being used to augment a critical workflow vs. just enable people to solve their own problems. Tools like excel work as horizontal technologies because the user of the tool can figure out how to leverage it to do their unique work. But when you’re hiring workers to do a job for you, you don’t also want to have to do all the work to train them on every single part of how to do their job. When you hire for work today, you rarely want it to be “generic”; instead you want it to solve a particular problem you have. The same is going to be true of AI Agents. You will want agents that are purpose built for the domains or problems you’re trying to solve. Which will lead to agents that are specialized by vertical or particular type of workflow. The game then is to build agents that combine a deep domain understand, propriety (customer) data that the models weren’t inherently trained on, a tight naturally interaction with the user’s workflow (either today’s or reimagined), and all of the resources necessary for the change management for the customer to be successful.
Aaron LevieCo-founder & CEO of Box
观点原文
当下 AI 最大的机会，很多来自那些过去从未有软件能够解决的问题领域，或那些规模小到不值得开发软件的细分市场。
AI Agent 的出现，是第一次真正实现了“软件 + 工作”一起被交付给用户。这意味着即便是在传统上软件支出很小的行业，TAM（可服务市场规模）也会被放大得非常巨大——例如法律服务、医疗、会计、软件工程等等。
关键在于：这些市场往往需要对领域问题的深刻理解，因为 AI 不是单纯帮用户“启用一个工具”，而是真正参与到某条关键工作流中，成为增强劳动力的一部分。
像 Excel 这样的工具之所以能成为通用型技术，是因为用户自己能想办法用它完成各自独特的任务。
但在现实中，当你雇人来替你做一项工作时，你并不想亲自从头到尾训练他如何做所有的细节。当你今天雇佣劳动力时，你几乎不会希望对方是“通用型”的；你希望对方能解决你特定的问题。
AI Agents 也会如此。
用户会希望获得为特定领域或特定问题量身打造的 Agent，这将推动 Agent 按垂直行业或按特定工作流高度专业化。
因此真正的“游戏规则”是：
打造既具备深度领域知识，又能利用专有（客户）数据，并能紧密嵌入用户工作流（无论是现有的还是重新设计的），同时还能提供所有实施与变革支持资源，使客户真正成功的 AI Agents。
Many of the biggest opportunities in AI right now are going after categories where there was never software to accomplish the problem before, or the categories were too small to matter.
AI agents represent the first time ever where we can bring along the work with the software, which means TAMs get very big even in spaces that traditionally had small software spend. Legal services, healthcare, accounting, software engineering, and so on.
The key though is a lot of these markets require a deep understanding of the problem in the domain because you’re fundamentally being used to augment a critical workflow vs. just enable people to solve their own problems.
Tools like excel work as horizontal technologies because the user of the tool can figure out how to leverage it to do their unique work.
But when you’re hiring workers to do a job for you, you don’t also want to have to do all the work to train them on every single part of how to do their job. When you hire for work today, you rarely want it to be “generic”; instead you want it to solve a particular problem you have.
The same is going to be true of AI Agents. You will want agents that are purpose built for the domains or problems you’re trying to solve. Which will lead to agents that are specialized by vertical or particular type of workflow.
The game then is to build agents that combine a deep domain understand, propriety (customer) data that the models weren’t inherently trained on, a tight naturally interaction with the user’s workflow (either today’s or reimagined), and all of the resources necessary for the change management for the customer to be successful.
AI
相关观点
- 21 世纪的技能 围绕“在 AI 智能体自动化大量当今任务的世界里，熟练工作者的未来会是什么样子”，网络上已经出现了大量讨论。 随着我们最近看到越来越多的例子——AI 编程智能体已经能够处理更长、更复杂的任务——一些人开始产生一种近乎存在主义的焦虑：在一个 AI 几乎什么都能做的世界里，人类以及我们现有技能的角色还剩下什么？ 甚至连 Claude Code 的创造者 Boris Cherny 都提到，在去年 12 月底的 30 天里，他贡献的代码 100% 都是由 Claude 写的。而我们也看到了无数工程师的案例：他们同时运行几十个 AI 智能体，每个负责项目的不同部分，在多个智能体之间快速切换，在几个小时或一两天内完成过去需要数周才能完成的工程工作量。 如果 AI 可以替我们完成所有编码工作，那问题自然来了：未来还需要熟练的工程师吗？更广泛地说，在 AI 能瞬间完成当今大量工作内容的情况下，那些拥有深厚领域知识的人——无论是法律、市场、建筑，还是医疗领域——还会有位置吗？ 答案是：毫无疑问，有。事实上，我甚至认为：历史上从未有过一个时代，比现在更适合成为某一领域的高度专业人才。 如果你仔细观察一名工程师今天的工作，就会发现它远不只是写代码那么简单。 工程师需要： 理解客户需求（或来自产品经理的需求） 测试代码 将代码整合进更大的系统 确保顺利发布 修复后续出现的 Bug 定义新功能 根据客户问题规划下一阶段功能 为安全和依赖升级各种库和包 以及几十项其他任务，最终交付一个真正可运行的软件系统 AI 正在逐步能够单独完成其中任何一项任务，但工程师的真正工作，是把这一切成功地整合在一起。 而那些真正理解这些任务如何协同运作的人，正是最有能力借助 AI 创造最大价值、获得最大回报的人。 这也是为什么你认识的很多顶尖工程师，正在 AI 编程时代玩得非常开心。因为他们不再需要把同样多的时间，花在“真正创造价值的工作”和“重复、枯燥的工作”上。现在，他们可以把那些苦力活交给智能体，释放时间，专注于更有趣、更有价值的部分。 历史不会简单重复，但往往会押韵。 你可能会想：这难道不会从根本上改变工作的本质，让人类能做的事情越来越少，最终让技能变得毫无意义吗？ 事实上，历史早已给出了答案。 如果你回顾任何一个高技能职业，从 个人电脑出现之前 到 AI 时代，几乎每一份工作今天的样子，都与几十年前完全不同。 在 1982 年的电影《大审判》中，有一个场景：Paul Newman 和 Jack Warden 饰演的角色，为了一场即将到来的审判，花几个小时翻阅一整柜法律书籍做研究。 光是看那个画面，你就能感受到当年这项工作有多么漫长、繁琐。 而仅仅几年之后，一名律师就可以通过电脑，几乎瞬间检索到同样的信息。 正如 Sam Altman 所说，一个自给自足的农民，可能会觉得今天的工作“轻松又琐碎”。同样，1982 年的律师，可能会认为如今的法律工作变得太简单，几乎不需要思考或研究。 但现实恰恰相反：法律行业随着监管复杂度、法律风险的提升，变得更加复杂和精密。法律的核心技能并没有消失，反而需求更高。一个反直觉但真实的数据是： 根据美国律师协会（ABA）的统计，活跃律师人数从 1975 年的约 40 万人，增长到 2025 年的约 137.5 万人。 再看看工程行业。就在 AI 编程真正普及之前，很多事情已经比几十年前简单得多了： * 开源库可以直接下载，解决几乎所有常见问题 * 工程师不再需要亲自管理基础设施 * 不用再自己处理数据库分片、身份系统、计费系统等 但与此同时： * 软件数量前所未有地增长 * 系统复杂度也达到了历史最高 结果是：我们对专家的需求并没有减少，反而更高了。 历史告诉我们的结论是：工作不会消失，只会变形。 当某些事情因为自动化或资源丰富而变得更容易，我们对“能做到什么”的期望就会随之提高，而技能的重要性，并不会下降。 工程师需要交付的内容范围更广了；律师被要求提供的判断和建议更复杂了；银行家和咨询顾问在表达、市场理解上的要求，也提高了数倍。 技能，作为杠杆 Max Levchin 最近说得非常好：“就像我们今天认为代码生成 AI 是 CS 学位的 10 倍生产力放大器一样，再过一两年，我们会反过来认为：CS 学位是代码生成 AI 的 10 倍放大器。” 你可以把“CS”和“代码生成”替换成任何一项技能，效果都是一样的。 AI 带来的真正杠杆，来自你在所自动化领域中的知识深度。 想要真正驾驭 AI 智能体，最重要的基础能力包括： 知道该把什么任务交给 AI 知道如何正确描述这些任务 能判断任务是否被正确执行 知道在 AI 给出结果之后，下一步该做什么 更关键的是，你必须能在脑中完整地理解： * 智能体正在做什么 * 接下来应该往哪里走 * 什么样的结果才算“好” 这种判断力，才是你能否持续、规模化创造价值的关键。否则，你只是在“乱点按钮”，而不是在真正使用 AI。 具体来看： 专业 CG 动画师，知道如何引导 AI 生成最好的画面，并把它们组合成能打动观众的情绪 专业律师，知道什么时候 AI 给出的建议在现实中不可行，或者因细微语境必须推翻 专业工程师，能判断智能体是否走进了未来无法扩展的架构死胡同 专业市场人员，知道哪种信息最能打动核心受众 未来，这些工作的“操作方式”都会彻底不同。你可能已经认不出专家在键盘上输入的内容，但他们理解“自己在做什么、为什么这么做”的能力，将一如既往地重要。 工作不会因为 AI 而被简单“替代”，而是会被不断扩展，被要求交付更高层次、更新类型的成果。 我们会： 对软件要求更高 对医疗突破期待更大 对医疗从业者的专业判断要求更强 几乎所有领域，门槛都会被抬高。 举一个小例子：在 Box，我们的解决方案工程师（销售团队中高度技术化的角色），现在可以为客户构建完全定制、可运行的 API Demo，并直接嵌入客户的软件系统中。 过去，这种展示对每一个客户来说成本极高，几乎不可能。但现在，它成了一件快速完成的事情。 而当这种能力在市场中普及后，客户自然会开始期待同样高度定制化的体验。 就在 AI 自动化了“演示”和“架构”的基础工作时，解决方案工程师的角色，反而被扩展成更具创意、更复杂的工作。 AI 的另一个惊人之处在于：它不仅放大了专家，也让普通人更快成为专家。 Dylan Field 将其形容为：AI 降低了地板，同时抬高了天花板。 如今，每一个有动力的人，都拥有一个 24/7 的全能导师。 我小时候学习做网站，纯属运气好，邻居（后来成为 Box 联合创始人）正好是程序员。而现在，每个人都有这样的“邻居”。 此刻，下一个 Zuckerberg 正在因为 AI 而第一次接触编程。下一个 George Lucas，正在用 AI 构建前所未见的世界，创造过去需要数千万美元才能实现的娱乐体验。 生命科学、金融，以及几乎所有技术领域，都会发生同样的事情。 如果你保持好奇、足够主动、真正渴望学习新技能，现在是人类历史上最好的时代。 在 21 世纪，技能不会因为 AI 而消失。它们所带来的杠杆，只会变得更大。 Skills in the 21st Century There’s been a lot of debate online around what the future of skilled work looks like in a world of AI agents automating much of the tasks that we do today. With all the examples we’ve seen most recently of AI coding agents being able to handle longer and more complex tasks, this has created a sort of existential dread in some around what the role of humans and existing skills will be in a world where AI can do almost anything. Even Boris Cherny, the creator of Claude Code, noted that 100% of his contributed code was written by Claude in the past 30 days in late December. And we’re seeing countless examples of engineers that are running up to dozens of agents in parallel, all handling different parts of a project, swiftly jumping between the work of each one, doing the equivalent of weeks of engineering work in hours or a couple of days. If AI can do all the coding for us, this begs the question, is there even a role for skilled engineers in the future? And more broadly, is there a role for anyone with deep domain expertise -whether it’s legal, marketing, architecture, or medical fields- in a future where AI can instantly do so much of what today’s jobs look like? The answer, emphatically, is yes. In fact, I’d argue, there has never been a time in history where it’s been better to be a highly skilled in a particular field. When you assess what an engineer does today, the process is far more complex than just writing code. You’re taking in requirements from the customer (or from a product manager, etc.), testing your code, integrating it with a broader project, ensuring that it ships properly, fixing future bugs and defining new features, handling issues as they arise, coming up with the next set of features to build based on customer problems, upgrading various libraries and packages for security issues and new dependencies, and dozens of other individual tasks that deliver real, working software. AI increasingly can do any single one of these tasks individually, but job of an engineer is putting all the work together, successfully. And those that understand how those tasks all come together are in the best position to deliver the most value, and get the greatest upside, with AI. This is why many of the best engineers you know are having the time of their life right now with AI coding. Because instead of having to spend a uniform amount of time on both the parts of coding that are value-add and differentiating from the stuff that’s rote, you can deploy tasks to agents that offload to the drudgery parts, freeing up more time for the work that is fun and delivers the real value. History doesn’t repeat itself, but it often rhymes Now, you may be wondering, this must fundamentally change the nature of work in a way that leaves less to be done by people, ultimately rendering skills useless, right? In fact, there’s already plenty of historical precedent for this evolution of work. If you look at the processes of any high-skill job starting from the pre-PC era to the AI era, nearly every single job would look fundamentally different today than it did just a few decades ago. There's a scene in the 1982 movie The Verdict where Paul Newman and Jack Warden's characters are combing through volumes of legal books for hours while doing research for an upcoming trial. You can just feel from watching the scene how long and tedious this process once was. Just a few years later, a lawyer would have pulled up a computer and retrieved the same data nearly instantly. Just as Sam Altman said a subsistence farmer might look at today’s jobs as way too easy and trivial, the lawyer from 1982 might imagine how easy law has become and devoid of critical thinking or research. And yet, the nature of the legal profession has only become more complex and sophisticated over time with a growing range of regulatory challenges, legal risks companies face, and so on. The foundational skills of the legal profession have been largely unchanged, and counterintuitively, despite many efficiency gains over time, the ABA pegs active attorneys having gone from roughly 400,000 in 1975 to roughly 1,375,000 in 2025. Or take that modern engineering job, right before the moment AI coding went mainstream. Many things have gotten simpler in the past few decades of coding: there are open source libraries you just pull down instantly for almost every problem some other engineer has once dealt with in the wild, and you no longer have to think about managing infrastructure, sharding databases, handling identity services, doing billing, or an array of tasks that every engineer had to do themselves just a couple decades back. And yet, we now have more software than ever before to manage and more complexity in our systems than ever before; and as a result, no less of a need for experts to help with all this work. What we know from history is that the jobs simply morph over time. We make one thing easier through automation or abundance, and as a result our expectations begin to grow on what we can now do in that space, and the skills become just as important. The expectations of what an engineer must deliver on today has just grown in scope, the demands of a lawyer’s advisory capabilities just increase, and the presentation and market knowledge you expect from your banker or consultant has gone up tenfold. Skills as leverage Max Levchin said it best just this week: “just as today we think of codegen AI as a 10x productivity multiplier for CS degrees (or equivalent), in a year or two we will be thinking of CS degrees as 10x productivity multipliers over codegen AI.” You can replace “CS” and “codegen” for any skill and see the same effect. The real leverage from AI comes from your depth of knowledge in the field you’re bringing automation to. One of the most important foundations for getting ahead with AI agents is being able to understand what tasks to give them, how to frame those tasks, how to know when those tasks aren’t being executed successfully, and then figuring out what to do *after* the AI has delivered the task. Critically, being able to hold a full mental model in your head of what the agent is doing, where it should go next, what “good” should look like is the critical judgment necessary to produce real value with agents vs. just use them to take shots in the dark. Without this underlying knowledge you’re going to have a very hard time producing ongoing value at any reasonable scale or frequency. The expert CG animator will know the right ways to get the AI to produce the best scenes, and then stitch them together the best to get emotion out of the audience; the expert lawyer knows when advice for a case would be impractical or when the AI should be overridden due to subtle nuance about a situation; the expert engineer knows when an agent has backed itself into an architectural dead-end that won’t scale for future features; and the expert marketer knows the message that will resonate with their core audience best. By today's standards, all of these jobs will look fundamentally different in the future. We'll barely recognize the things that the experts are typing into their machines to do their work, but all of their underlying knowledge and skills of what they are doing and why will be as necessary as ever. And instead of people and companies simply using AI agents to automate the work that we already do manually, jobs will simply be expanded over time to expect higher, or new types of output. We’ll demand far more from our software, we’ll expect more medical breakthroughs, require better expertise from our healthcare providers, and bar will just rise in essentially every other domain out there. As a minor example of this dynamic, at Box, we now have solutions engineers (the deeply technical counterparts to our sales team) build fully custom, working demos of Box’s APIs embedded into our customer’s software. Instead of a static, generic demo, they can now bring the full platform to life in a way that would have been cost prohibitive to demonstrate for every client before. But now it’s a quick task. As this trend propagates across the market, customers will soon expect similarly customized experiences when they’re pitched by software companies. Just at the same moment when the rote task of pitches or architectures gets automated by AI, a new set of more creative and elaborate tasks become required as a part of a solution engineer’s job. And here’s the amazing thing with AI. Just as it makes the experts far more capable than before, it also makes it so anyone can learn to become an expert much faster. Dylan Field frames this as AI lowering the floor and raising the ceiling. In his case, it means anyone who’s motivated can now learn how to do amazing design if they want. The tools that every young person now has available to them is unprecedented. Anyone, if they so choose, has a complete tutor available to them in every skill on the planet. When I was growing up and getting into building websites, it was only by sheer coincidence that my neighbor (and ultimately a Box cofounder) happened to be a programmer that could help multiply my skills; but now everyone has that neighbor available to them 24/7. Right now, the next Zuckerberg is getting into coding for the first time because of AI. The next George Lucas is someone developing AI worlds that have never been seen before, and producing a level of entertainment that would not have been possible even a couple of years ago without tens of millions of dollars. The same will be true in life sciences, finance, and every other technical field out there. This is the greatest time in history if you’re curious, resourceful, and are fundamentally want to learn and pick up new skills. Skills don’t go away in the 21st century with AI. The leverage you get from them have only gone up.Aaron Levie · 2026/01/11
- 当下，对任何具备创业精神的人来说，在自己所在公司里最大的机会之一，就是成为那个帮助组织理解并部署 AI 智能体（AI agents）的人。 部署智能体并不容易。目前真正掌握这项能力的人仍然非常有限。虽然几乎所有最佳实践都能在网上找到，但前提是——你必须持续、高强度地关注这个领域的每一个新进展：包括编程型智能体、记忆机制、文件系统与工具调用、Agent 框架（agent harnesses）、上下文工程（context engineering）等等。 更重要的是，这个领域的最佳实践会不断复利式叠加。这意味着：你领先得越早，就越容易理解并部署下一波新出现的能力。 如果你能做到这一点，那么在你加入的任何一家公司里，你都会看起来像是拥有一台时间机器的人。 这是一个巨大的机会。 One of the biggest opportunities available to anyone entrepreneurial right now in their company is to be the person that helps their organization understand and deploy AI agents. Deploying agents is hard. There is still a relatively narrow universe of people with this skill. While all of the best practices are available online, you have to be paying attention to every new thing that comes out around coding agents, memory, filesystem and tool use, agent harnesses, context engineering, and so on. And the best practices in this space just continue to compound, which means the farther ahead you are the more you understand about how to deploy what comes up next. So if you can do that, you will look like you have a time machine to any company you join. Huge opportunity.Aaron Levie · 2026/01/07
- 上下文时代 彼得·德鲁克（Peter Drucker）——管理与商业领域最重要的思想家之一（也是“知识工作者”一词的提出者）——在 1990 年代就指出：“知识已经成为关键的经济资源，并且是占主导地位、甚至可能是唯一的竞争优势来源。”他对 AI 时代的预判，几乎精准到不可思议。 随着 AI 进步持续不减，AI 模型将越来越能够增强或自动化各种知识工作者的任务。正如我们在编程领域看到的那样，这些模型将演化为 AI 代理：在几乎所有工作领域中，成为专家级律师、医疗专业人士、商业战略家、科学研究者等。 但默认情况下，这些模型并不了解你的具体团队或组织。它们可能瞬间被要求为 福特审查法律合同，下一秒又为 高盛编写新软件代码。它们是完全通用的超级智能系统，可以为任何提出请求的人执行任何任务。其结果是：你的公司与另一家公司拿到的是同一个“专家律师”、同一个“工程师”。 因此，我们必须面对的问题是：在人人都能获得同等智能的世界里，公司如何实现差异化？ 当然，差异化的一部分来自团队和员工如何高效使用 AI 代理；但真正的终极放大器将是代理所获得的上下文（context）。这些上下文包括：要打造的正确产品、服务客户的方式、要进入的市场、客户互动的具体细节、组织内部的“部落知识”，以及随着时间积累、授权或并购而形成的、只属于你公司的大量专有信息。 过去一年，“上下文工程（context engineering）”为解决这一问题而迅速发展，但它并不容易。想象一下：一个默认对你组织一无所知的专家律师或工程师，你却只有一份文档的空间来向他说明其全部工作内容、可使用的系统、所需数据、目标等等。把正确的信息交到代理手中，成为提升生产力的关键前提。 事实上，这个问题在 AI 代理出现之前就困扰着企业。公司一直是各种上下文的集合体，试图从中榨取最大价值：流程、知识产权、独特想法与路线图、决策方式、客户信息等等。然而，随着规模扩大，企业只能利用其中很小一部分知识（无论是否已数字化）。惠普（HP） 90 年代的 CEO 卢·普拉特（Lew Platt） 曾说：“如果 HP 知道 HP 所知道的一切，我们的生产力将提升三倍。” AI 代理终于让这一愿景成为可能。我们第一次能够真正动用沉睡在组织内部的知识与信息。Atlassian 的 迈克·坎农-布鲁克斯（Mike Cannon-Brookes） 将其形容为一种“数据的梅特卡夫定律”：就像网络效应一样，数据越多，整体系统越强。 这意味着，在 21 世纪，最关键的竞争优势之一，将是企业捕获、管理并围绕正确上下文构建流程的能力。 拥有更好市场定价与租金洞察的房地产公司会赢得更多客户；能利用海量数据开发新药的制药公司会创造更多收入；能产出更好营销活动的代理机构会更具差异化。 然而，把正确的数据（上下文）交给代理，在规模化时是个难题。正如 Foundation Capital 的 贾娅·古普塔（Jaya Gupta） 和 阿舒·加格（Ashu Garg） 在其关于上下文图谱的犀利文章中所指出的那样，许多 AI 代理所需的关键决策轨迹，并不存在于任何现有软件中。而这只是代理所需上下文的一种：它们还需要客户数据、企业文档、对话历史、项目时间线、研究数据、营销素材、财务记录、代码库、HR 信息等等。 因此，设计系统以让代理访问这些数据，并确保所有代理能在这些数据上互操作，至关重要。在一些情况下，现有的系统将成为这些数据与代理的合理“归宿”；而在另一些情况下——尤其是当今天还不存在相应软件时——我们将需要全新的工具。即便如此，未来十年里，数据访问控制与权限冲突、治理与合规、以及连接分散的企业系统，都将是持续的工作。 此外，企业还必须推进大量变革管理。我们曾以为 AI 会适应我们的工作方式，但事实证明，鉴于其强大的能力与内在限制，我们反而需要去适应 AI 的工作方式。这意味着必须优化组织与工作流，以便为代理提供成功所需的上下文。其核心原则是：用户现在要负责指挥和引导代理的工作，并确保在过程中提供正确的上下文。 本质上，当下的个人贡献者，将在未来成为“代理的管理者”。他们的新职责包括：监督与升级路径、跨代理的协调，以及在不同代理之间“护送”工作——就像前 AI 时代的团队管理者一样。 总体趋势已经十分清晰：我们正进入“上下文时代”。 能够积累并最好地利用上下文的团队与公司，将实现最高的生产力与产出；做不到的，将越来越难以具有竞争力地服务客户。正如彼得·德鲁克当年所预言的那样。 The Era of Context Peter Drucker, one of the most important thinkers on management and business (and coiner of the term knowledge worker), in the 90s said, "knowledge has become the key economic resource and the dominant-and perhaps even the only-source of competitive advantage." He couldn’t have predicted the AI era more accurately. As AI progress continues unabated, AI models will get more and more capable of augmenting or automating a variety of knowledge worker tasks. Just as we've seen in coding, these models will morph into AI agents that are expert lawyers, healthcare professionals, business strategists, scientific researchers, or other roles in almost every domain of work. But these models, by default, don't know anything about your particular team or organization. In an instant, they might be asked to review a legal contract for Ford, and then in the next second be writing code for a new piece of software at Goldman Sachs. They are fully general-purpose superintelligence systems that can take on any task for anyone that’s asking. As a consequence, that means that your company is getting the same expert lawyer as another company, the same engineer, and so on. The question that we will have to wrestle with is, in a world where everyone has access to the same intelligence, how does a company differentiate? Certainly it will be about how teams and employees use AI agents effectively, but the ultimate force-multiplier will be the context that the agents get. Context about the right products to build, the ways to serve customers, the markets to go after, the specific details of client interactions, the tribal knowledge in an organization, and endless amounts of other proprietary information to your company that's developed over time, licensed, or acquired. Context engineering has blown up over the past year to tackle just this, and it's not an easy problem. Just imagine taking an expert lawyer or engineer that by default knows absolutely nothing about your organization, and you only have a single document’s worth of space to describes their entire job, every system they have to leverage, all the data that they are supposed to work with for that particular task, what their objectives are, and so on. Getting the right information to them becomes a critically important requirement to drive their productivity. Now, this problem has already plagued companies well before agents existed. Companies have always been a collection of various forms of context that they try and extract the most value out of: their processes, intellectual property, unique ideas and roadmap, ways they makes decisions, information about their customers, and so on. And companies, as they scale, have only been able to utilize a small portion of this knowledge (digitized or not). Lew Platt, the CEO of HP in the 90s, once said, “if HP knew what HP knows, we would be three times more productive.” AI Agents finally make this vision possible. We can, for the first time, begin to tap into this wealth of knowledge and information sitting inside of our organizations. Mike Cannon-Brookes of Atlassian once framed this as a sort of a Metcalfe's law of data. Like the original theory on network effects, the more data you have, the more powerful the overall system becomes. This means that in the 21st century, one of the most critical forms of competitive advantage will be a company's ability to capture, manage, and build processes around the right context. The real estate firm that has better insights on market pricing and rent dynamics will get more clients, the pharmaceutical company that can develop and design new drugs using reams of data will generate more revenue, the marketing agency that can generate better campaigns will differentiate more. Now, getting the right data (or context) to agents is a hard problem at scale. As Jaya Gupta and Ashu Garg at Foundation Capital pointed out in their provocative essay on context graphs, many of the critical decision traces for AI agents to operate with either aren't in any existing software today. And this is just one example of the kind of context that agents will need to operate on. They’ll need customer data, enterprise documents, conversation history, project timelines, research data, marketing assets, financial records, codebases, HR information, and much more. Designing our systems to get agents access to that data, and ensuring that all of our agents can interoperate on that data is going to be incredibly important. In some cases existing systems of record will become the logical houses for that data and the corresponding agents; and in other cases -especially where there is no such software today- we will need all new tools, just as Jaya and Ashu laid out. Even then, managing the myriad conflicts that emerge in data access controls and permissions so agents don’t leak data, governance and compliance challenges around what things an agent can execute on, and connecting our disparate enterprise systems to make this work is going to be an ongoing work for the next decade. Further, companies will have to drive a substantial amount of change management to make this all work. We imagined that AI systems would adapt to how we work, but it turns out due to their extreme power (and inherent limitations) we will instead adapt to how they work. This means we will have to optimize our organizations and workflows to best enable context for agents to be successful. The core tenant of this change is that the user is responsible now for directing and guiding agents on how to do their work, ensuring it gets the right context along the way. In essence, the individual contributor of today becomes the manager of agents in the future. Their new responsibilities will be providing the oversight and escalation paths, a meaningful amount of coordination throughout the work that the agents are doing, and shepherding work between the various agents - just like managers of teams in the pre-AI era. The overall trend is clear, and we are entering the era of context. The teams and companies that can accumulate and best utilize context will drive the greatest productivity and highest output. Those that don’t will find it harder and harder to serve customers competitively. Just as Peter Drucker predicted.Aaron Levie · 2026/01/05
- 这就是为什么上下文工程（context engineering）如此重要。 我们刚刚花了 2 个小时，仅仅围绕一种问题类型，讨论一个智能体究竟应该依赖自身的内部知识，还是尝试在数据中寻找相关上下文。而在数百个测试用例中，我们只讨论完了 2 个。 甚至连参与这次头脑风暴的人，也无法就“在这种情况下，人类会怎么做”达成一致。这里根本不存在唯一正确的答案，而且答案始终是高度依赖具体情境、并且因客户而异的。 上下文工程中的一切，都是在多个维度之间做取舍： 你希望智能体多快给出答案？ 你希望用户需要多少来回交互？ 在尝试回答问题之前，它应该先做多少准备工作？ 它如何判断自己已经拥有足够完整的来源材料来回答问题？ 给出错误答案的风险有多高？等等。 你在其中任何一个维度上做出的决定，都会在另一个维度上产生后果。没有免费的午餐。这正是为什么构建 AI 智能体如此疯狂、也如此困难。 这也凸显了一个事实：在 LLM 之上，仍然蕴藏着巨大的价值空间。把这些决策做对，直接决定了产品价值主张的质量。 Here’s why context engineering is such a big deal. We just spent 2 hours debating when an agent should rely on its internal knowledge vs. trying to find relevant context within data for just one type of question. We got through 2 test cases of hundreds. Even the people involved in the brainstorm couldn’t all agree on what they would expect humans to do in this situation. There truly was no right answer, and it’s always context specific customer by customer. Everything in context engineering is a tradeoff between a variety of factors: how fast do you want the agent to answer a question, how much back and forth interaction do you want to require for the user, how much work should it do before trying to answer a question, how does it know it has the exhaustive source material to answer the question, what’s the risk level of the wrong answer, and so on. Every decision you make on one of these dimensions has a consequence on the other end. There’s no free lunch. This is why building AI agents is so wild. It also highlights how much value there is above the LLM layer. Getting these decisions right directly relates to the quality of the value proposition.Aaron Levie · 2025/12/24