Sequoia Capital Declares Shift to Computational Revolution as AGI Arrives via $10 Trillion AI Agent Market

红杉资本：计算革命已至，AGI将依托10万亿美元AI智能体市场到来 该风投机构阐释了具备长周期执行能力的自主智能体如何通过将人工认知工作转化为可无限扩展的Token化服务，从而把长达百年的开发周期压缩至100天。 本次主题演讲概述了在转向“计算革命”的根本驱动下，行业正经历从传统软件向规模达10万亿美元的AI服务市场的结构性转变。演讲者提出了初创公司在模型能力飞速迭代中生存发展的“MAD”策略（护城河 Moats、功能可供性 Affordance、扩散 Diffusion），并梳理了自主AI智能体从简单助手演进为独立“黑灯工厂”的历程。通过与工业革命作类比，演讲预测未来机器将承担99.9%的认知工作，推动人类精力转向纯粹的创造力、战略思考以及本质的人际连接。 红杉资本指出，当前 AI 浪潮标志着从“信息分发的通信革命”向“信息处理的计算革命”的根本转变，并将解锁规模达 10 万亿美元的全球服务市场（TAM），体量远超过去的云计算转型。 (1:56) 从商业与应用实践角度来看，能够自主从失败中恢复并长周期执行复杂任务的 Agent 出现，标志着 AGI 在实用层面已经正式到来。 (4:13) 面对底层模型能力的快速迭代，初创团队需借助 MAD 框架（护城河 Moats、易用性 Affordance、扩散 Diffusion），以自上而下的客户深度绑定与最低摩擦的产品体验构建持久壁垒。 (6:24) “服务即新软件”时代已经开启，随着以 Token 结算的自主 Agent 在无人工介入的场景下普及，研发与生产周期被急剧压缩，未来百年的构建工作有望在 100 天内落地。 (19:43)

Sequoia Capital partners argue that the rise of long-horizon AI agents marks the functional arrival of artificial general intelligence, driving a transition from a software market into a ten-trillion-dollar services economy. The Services Opportunity and Practical AGI (1:56) Pat frames artificial intelligence as a fundamental shift from a revolution in communication to one in computation. Whereas the cloud software transition expanded its total addressable market from $350 billion to $650 billion over 15 years, AI unlocks the global services market, estimated at approximately $10 trillion. To illustrate the magnitude, the speakers note that the United States legal services vertical alone represents a $400 billion market, matching the entire historical size of cloud software. The technological foundation has evolved across three key inflection points: the November 2022 ChatGPT launch demonstrating pre-training scale, the o1 reasoning phase establishing inference-time compute scaling, and recent long-horizon developer agents like Claude Code. While acknowledging they are venture investors rather than technical researchers, the speakers assert that systems capable of sustained autonomous execution, error recovery, and task completion fulfill the functional, commercial definition of AGI. Navigating the Technology Torrent With MAD (6:24) Because underlying model capabilities are shifting rapidly, Pat advises founders to approach development 'customer-back' rather than 'tech-out.' Under the MAD framework—Moats, Affordance, and Diffusion—durable enterprise moats come from deeply embedded customer workflows and relationships, which change far more slowly than model architectures. The framework defines affordance as building intuitive paths of least resistance, noting that complex terminal interfaces intimidate typical corporate users despite their technical power. Meanwhile, the diffusion gap between frontier model capabilities and enterprise IT adoption is actively widening. Using a racing analogy, the speakers suggest that rapid technological turbulence allows late entrants to overtake incumbents who cannot adapt quickly. Agent Architecture and Autonomous Dark Factories (11:51) Sonia tracks the evolution of AI agents from early, unreliable 2022 experiments like AutoGPT and BabyAGI to modern systems anchored by three pillars: reasoning and planning, tool usage, and harness-driven reinforcement learning feedback loops. Model stability on complex tasks has expanded from tens of minutes to sustained hours. Rather than eliminating traditional software, the speakers argue that millions of deployed agents will drastically increase utilization of existing software-as-a-service platforms. Autonomy is progressing along an operational spectrum: from early inline autocomplete tools, to human-managed agents, to self-managing agents, and ultimately to 'dark factories' where agents operate asynchronously and deploy directly to production without human review. The speakers contend that 'services is the new software,' pointing out that token-based agents are infinitely scalable and significantly cheaper than salaried employees across disciplines such as law, medicine, finance, and software maintenance. Historical Parallels and Industrial-Scale Cognition (21:03) Constantine draws a direct parallel between the Industrial Revolution and modern machine cognition. Historically, human and animal muscle accounted for almost all physical labor before mechanical engines scaled to perform more than 99% of physical work on Earth. The presentation forecasts a similar trajectory for knowledge work, projecting that machines will eventually handle 99.9% of all cognitive processing. To contextualize this shift, the speakers present historical analogies: aluminum transitioning from a rare, precious metal to cheap household foil through electrolysis; evolutionary computing producing alien, organic NASA antenna designs that outperform human symmetrical designs; and early steam engine tinkering preceding formal thermodynamics by over a century. They suggest current AI development remains in an empirical tinkering stage ahead of a formal scientific framework. The presentation concludes that while ubiquitous machine cognition will compress development cycles and automate cognitive labor at an unprecedented scale, intelligence itself possesses no intrinsic value without human utility; long-term commercial durability and cultural relevance will remain anchored in human experience and relationships.

红杉资本（Sequoia Capital）在AI Ascent 2026演讲中指出，具备自主纠错与长程执行能力的AI Agent标志着AGI在商业实践层面的落地，推动AI由信息分发的通信革命转变为信息处理的计算革命，并将价值捕获空间从数百亿美元的传统软件市场拓展至规模达10万亿美元的专业服务领域。 软件到服务：AGI商业落地与万亿市场 (1:56) 红杉资本合伙人Pat指出，当前AI浪潮与以往计算转型的核心差异在于其同时覆盖软件与服务。云计算在15年间推动软件总体有效市场（TAM）从3500亿美元扩张至6500亿美元（其中云占据4000亿美元），而AI则直接切入规模预估达10万亿美元的全球服务市场。以美国法律服务为例，该单一垂直行业规模即达4000亿美元，已等同于传统云软件的整体市场体量。驱动这一转变的根本原因在于AI并非类似互联网或移动设备的“通信革命”（改变信息分发），而是一场重构信息处理方式的“计算革命”。 从商业实践角度，红杉主张AGI已经实质到来。演讲将技术路径划分为三个拐点：2022年11月ChatGPT验证预训练潜能、o1模型开启推理期算力扩展规律，以及Claude Code展示的长程Agent执行力。红杉提出，当系统能够接受任务指派、遭遇失败自主恢复并持续推进直至目标达成，在实用意义上即符合AGI特征。相较于以往提供10%至40%效率微调的过渡产品，长程Agent正在带来10倍至40倍的生产力跃升，推动组织结构根本性重组。 创业防御机制：MAD框架与客户壁垒 (6:24) 面对底层模型能力的快速迭代，红杉面向初创企业提出了“MAD”战略框架，涵盖护城河（Moats）、示能（Affordance）与扩散（Diffusion）。在护城河构建上，演讲警告创业者切忌单纯由技术自内向外推演（Tech-Out），而必须自客户需求向后反推（Customer-Back）。底层模型能力虽日新月异，但企业客户的核心业务流程与组织关系演进相对缓慢；因此，将业务深度包裹于客户关系之中，是抵御技术底座变动的核心防御机制。 在示能与扩散维度，模型能力的复杂性与终端用户体验之间存在巨大断层。以Claude Code为例，终端命令行对多数大型企业员工门槛过高，初创企业的机遇在于构建极简直观的交互路径（如幼儿拿到锤子即知敲击）。此外，基础模型的进化速率已显著拉大与企业IT采纳周期之间的扩散鸿沟。红杉以“雨中赛车”作喻，认为当前频繁的技术涌现打破了原有竞争壁垒，既有领先者并不安全，任何敏捷的构建者均具备超越机会。 Agent演进路径：黑灯工厂与代币经济 (17:08) 红杉合伙人Sonia梳理了Agent的技术架构与自主权演进。Agent由模型（大脑，复杂任务执行时长已从数十分钟扩展至数小时）、工具（四肢，涵盖终端、Slack及SaaS API）与线束（Harness，提供专业领域的强化学习反馈闭环）三大支柱组成。自主权层级正在从辅助自动补全、需人工管理的实习生、自主管理的实习生，迈向完全剥离人工审查直接进入生产环境的“黑灯工厂”（Dark Factories）模式，该模式已在网络安全等容错率极低的领域落地运行。 在经济学层面，“服务即新软件”重构了用工逻辑。相较于招聘成本高、扩容困难且需支付固定薪资的人类劳动力，Agent具备依托算力无限扩展、维护成本低以及按代币（Tokens）计费的特性，边际执行成本大幅降低。长程Agent正在急剧压缩开发周期：例如Zed开发者单人利用Agent在假期完成三年期项目、Sierra创始人周末重构平台、Notion团队在六周内重写800万行代码。红杉预测未来将出现数以亿计Agent互联并直接进行交易结算的指数级增长阶段。 认知革命隐喻：智能过剩与人类价值锚点 (21:03) 演讲将认知革命与工业革命进行深度对照：在物理劳动领域，数百年机械化使机器承担了当前地球上99%以上的体力工作；而在思维领域，计算技术的飞跃预计将使未来地球上99.9%的认知任务交由机器处理。红杉通过多重视角阐释该趋势：铝在电解法诞生前曾贵如珠宝，技术突破后迅速变为廉价包材，智能亦将从高门槛技能演变为唾手可得的大宗消耗品；演化算法设计的NASA卫星天线展示了非直观但极高产出的“异星设计”范式；当前AI对数万亿代币的调试则类似于蒸汽机时代的工匠摸索，未来必将催生如热力学般的基础科学体系。 最后，演讲援引早期摄影技术对绘画艺术的冲击指出，具象写实的机械化促成了印象派、表现主义与立体主义的繁荣。借用古希腊哲学家普罗泰戈拉“人是万物的尺度”，红杉强调无论智能与劳作如何被机械化接管，事物本身的意义完全依赖于人类体验而存在。AI能够承担具体交付工作，但商业与社会协作的核心价值纽带依然植根于人与人之间的真实连接。 红杉资本认为，由长程Agent驱动的认知革命正以远超工业革命的速度重构商业与生产力格局。尽管Agent具备替代大量认知工作与压缩开发周期的潜力，但技术扩散的断层仍存；在基础模型日趋商品化、智能供给过剩的背景下，商业持久壁垒依然依赖于深度的客户场景绑定，且所有效率工具的最终评判尺度依然在人类体验本身。

[4:13] if you can dispatch an agent to do a job and it can recover from failure and persist until that job is done... That feels pretty much like AGI. [7:57] Pat: Your customers are not changing nearly as fast as the capabilities are changing [9:33] Pat: You cannot pass 15 cars in the sun, but you can pass 15 cars in the rain. [11:51] Sonia: cute, kind of endearing, but completely useless [11:51] Sonia: will literally snitch on your neighbors for tax fraud [13:45] Sonia: a system that perceives its environment, chooses actions, and progresses autonomously towards a goal. [13:45] Sonia: simply the ability to get done. [18:01] services is the new software. [19:43] whatever you could imagine building over the next 100 years, we think is now possible in 100 days. [21:03] 99 plus percent of all the physical work done on planet Earth for humans is done by machine. [24:45] aluminum is intelligence [24:45] electrolysis is artificial intelligence. [27:50] Protagoras: Man is the measure of all things. [27:50] only the human connection can give you a reason to care.

[0:02] Good morning. How's everybody doing? [0:04] >> [cheering] [0:05] >> All right. All right, a little bit [0:06] better. Hey, thank you all for being [0:07] here. We really appreciate it. We do [0:09] this as a service to the community [0:12] because we are living through important [0:14] times. [0:15] And it's an honor for us to be able to [0:17] serve as a bit of a gathering place for [0:19] people to come together. And this is by [0:21] far the best agenda we put together and [0:23] by far the best set of attendees. And so [0:25] we just want to start out by saying [0:26] thank you. We know you are all crazy [0:28] busy. Thank you for being here today. [0:32] >> [applause] [0:35] >> So, we do have a pretty exceptional [0:37] agenda put together as per usual. [0:40] Sonia [0:41] in the impeccable outfit [0:44] Constantine and [snorts] I are going to [0:46] say a couple of words to start. You [0:48] know, we have we have the privilege of [0:49] being in the midst of a lot of [0:51] conversations with a lot of interesting [0:52] people. And so once a year we like to [0:54] try to synthesize that a little bit and [0:56] share back with you what it is we've [0:57] been hearing. And so I'll say a few [0:59] words of overall [1:01] calibration and then Sonia will say a [1:03] bit about what we see today and then [1:05] Constantine will say a bit about what we [1:08] think might be coming tomorrow. [1:10] So, for calibration [1:13] we're going to start by zooming out. [1:15] Going back to the silicon [snorts] based [1:17] transistors which gave this area its [1:18] name. [1:19] They got built into systems connected by [1:21] networks that went public in the form of [1:23] the internet, supported applications [1:25] like social media and the cloud, [1:27] eventually showed up on our pockets in [1:28] mobile devices that today [1:30] are capable of doing something [1:32] indistinguishable from magic which is [1:34] AI. [1:36] The reason we like to show this slide [1:38] and those of you who've been here in the [1:39] past have seen this before [1:41] is because it reminds us [1:43] that all of these waves are additive. [1:46] And we sort of needed all of these [1:48] decades of evolution to have the [1:49] compute, the bandwidth, the data, the [1:52] talent to make the most of this moment. [1:56] Now this AI wave is a little bit [1:57] different in three ways. First [2:01] it's the biggest wave yet. [2:04] And that's generally true, but there is [2:05] something more specifically true about [2:07] this wave. [2:08] Which is it is the first one that is [2:10] both software and services. [2:13] The top row shows the first 15 years of [2:16] the cloud transition where the TAM for [2:17] software [2:19] went from about 350 billion to 650 [2:21] billion and cloud grew to be about 400 [2:24] billion of that. [2:25] The bottom row is what is brand new. [2:28] This is the services revenue [2:31] that seems to also be available now. [2:34] $10 trillion is a conveniently round [2:36] number. We don't know if it's 10 [2:37] trillion or 5 trillion or 50 trillion. [2:40] We do know that legal services in the US [2:41] alone is a $400 market. [2:44] That is one vertical [2:45] and one geo. [2:47] And it's the same as all of software. [2:50] So this opportunity is immense. [2:53] Point number two [2:55] fastest wave yet. [2:57] I think we can all feel this. [3:00] What it means [3:01] is that this white space, and I direct [3:03] your attention to the AI side of this [3:05] page this white space [3:07] is getting filled pretty fast. [3:10] These logos [3:12] are the companies that got to a billion [3:13] plus of revenue as a result of the cloud [3:16] mobile and now AI tectonic shifts. [3:19] And at current course and speed [3:22] there are more coming soon. [3:25] Point number three [3:27] which is probably the most interesting [3:28] one and I borrowed this [3:30] from my partner Constantine [3:33] is that there are two basic kinds of [3:34] revolutions in technology. [3:37] There are revolutions of communication [3:40] which are about the way information is [3:41] distributed. [3:44] Most of the people in this room have [3:46] only lived through revolutions in [3:48] communication. [3:50] The internet [3:52] the cloud [3:53] mobile [3:55] those are all about information [3:56] distribution. Those were revolutions in [3:58] communication. [4:00] AI is different. [4:01] AI is this one. [4:04] AI is a revolution in computation. It's [4:05] about how information is processed. [4:08] And that might sound like semantics, but [4:10] these are fundamentally different shapes [4:12] of waves. [4:13] And maybe the most visceral way to feel [4:15] this [4:17] is to think about the fact that the [4:18] floor keeps moving underfoot. The [4:19] technology foundation [4:21] on which everybody builds [4:24] changes every day when new capabilities [4:26] come out. [4:28] And we've had three major inflection [4:29] points [4:30] over the last handful of years. First [4:31] one [4:32] chat GPT chat GPT moment, November 2022, [4:36] the world saw the power of pre-training. [4:39] Second one, couple years later [4:41] 01 model reasoning. [4:44] All of a sudden a second scaling law [4:45] emerges around inference time compute. [4:48] Third one [4:50] just recently [4:51] cloud code no plus 45 now 47. [4:54] The world saw the power of long horizon [4:56] agents. [4:57] And while these look like three points [4:59] on a continuum [5:01] it's kind of a heartbreak between two [5:02] and three. It's a little bit of a [5:04] discontinuous change. [5:06] And if we may may be so bold [5:09] we would say [5:11] that this is AGI. [5:14] And look [5:15] I'm an econ major [5:16] we're venture capitalists [5:18] not about to propose a technical [5:19] definition for AGI. Okay, we study [5:22] founders and markets and the collision [5:24] thereof which is businesses. [5:26] But we do study businesses. And so from [5:29] a commercial standpoint, from a [5:31] practical standpoint from a functional [5:33] standpoint if you can dispatch an agent [5:37] to do a job and it can recover from [5:39] failure and persist until that job is [5:41] done [5:43] I don't know. That feels pretty much [5:45] like AGI. [5:47] Even if you don't think it's AGI, which [5:49] is fine. [5:50] By the way, Sonia will talk a lot more [5:51] about this in her part. [5:53] Even if you don't think it's AGI [5:55] I think we can all see that the cars [5:57] have arrived. [5:58] Last few years we've had a lot of faster [6:01] horses. [6:02] Applications that made you 10 or 40% [6:04] more productive, but didn't [6:06] fundamentally change the way you work. [6:09] Now we're starting to see cars. [6:11] Applications that make you 10 or 40X [6:13] more productive. [6:14] And absolutely change the way that you [6:16] work. Change the nature of your work. [6:18] Change the nature of your organization. [6:20] Cars have arrived. [6:24] This is the founder of Sequoia, Don [6:25] Valentine. He was famous [6:27] for asking one question. [6:30] So what? [6:31] Why does all this stuff matter? [6:33] >> [clears throat] [6:34] >> Well, it matters cuz just in the last [6:35] few months [6:37] the race has begun. [6:40] And it's a different kind of race [6:42] than what we're used to. [6:44] The way you drive a car is different [6:45] than the way you ride a horse. The way [6:46] you build a car is different than the [6:47] way you take care of a horse. [6:49] So it's a very different sort of race. [6:51] And one of the reasons that we wanted to [6:52] gather everybody here today is because [6:54] nobody has all the answers. [6:56] And the more time we can spend together, [6:57] the more we can learn and hopefully [6:59] figure out [7:00] where all this stuff is headed. [7:02] And it's important that we do so as soon [7:04] as possible [7:06] cuz there's a lot at stake. [7:08] Just from a commercial perspective [7:10] there's $10 trillion up for grabs. We [7:12] got labs coming at it from a tech out [7:14] approach. We got startups building on [7:16] top coming at it from more of a customer [7:17] back approach. [7:19] We do have all of the labs represented [7:21] in this room, but most of you are [7:23] building on top. And so we'll take a [7:24] minute and uh talk about that customer [7:27] back approach. [7:29] So our advice for those of you [7:31] who are building on top of the labs [7:33] is free advice and so it's worth every [7:35] penny you paid for it. Our our advice [7:38] would be to get mad. [7:41] And we don't we don't actually need you [7:42] to be angry. You can be angry if you [7:43] want. If that's what drives you, that's [7:44] cool. Go ahead and be angry. But this is [7:46] just a convenient acronym [7:48] for motes, affordance, and diffusion [7:52] which are three characteristics or three [7:53] pillars [7:55] of a strategy for building on top of the [7:57] models. First on motes [7:59] just for fun. Anybody remember this [8:00] slide from last year? [8:03] One person and he's my partner. Okay, [8:05] that's cool. Um [8:07] we're we're not going to [8:08] Okay, as a reminder this slide shows the [8:10] merchandising cycle which is the links [8:12] in the value chain required to take [8:13] something from an idea to a happy [8:15] customer. We're not actually going to go [8:17] through the links in the chain. [8:19] The point that I want to make here [8:21] if you approach things from a tech out [8:22] point of view, each link in the chain [8:24] gets approached a little bit different. [8:26] If you approach things from a customer [8:27] back point of view, each link in the [8:29] chain you approach a little bit [8:31] different. [8:32] Now here's the part that's [8:33] counterintuitive. [8:35] In a revolution of computation [8:38] which is about information processing [8:41] what you want to do [8:43] is look down here [8:45] because there's cool new stuff coming [8:46] out all the time. [8:48] What you should actually do for the sake [8:50] of building motes [8:52] is look up here [8:54] because your customers are not changing [8:56] nearly as fast [8:58] as the capabilities are changing. [9:01] The things that you built might be [9:02] irrelevant tomorrow. [9:05] The degree to which you wrap yourself [9:06] around your customers is going to be a [9:08] bit more durable. [9:10] That's not to say the product and [9:11] technology is not important. It is [9:13] insanely important. And generally [9:15] speaking best product wins. [9:17] But in a world where product changes so [9:20] fast because capabilities change so fast [9:23] in thinking about motes, we would [9:25] encourage you to go as customer back as [9:27] possible and think about all the ways [9:29] you can wrap yourself around those [9:31] customers. [9:33] Okay, the A in MAD stands for [9:35] affordance. [9:36] This is a term that we borrow from the [9:37] design world. [9:39] A hammer [9:40] is an object that has affordance. [9:42] I have a two-year-old son. [9:44] If I give him a hammer [9:46] he would know what to do with it. He [9:47] would grab it and start hitting stuff. [9:48] That's why we don't give him hammers. [9:50] Okay? An object with affordance [9:52] is one that doesn't need to be [9:53] explained. [9:55] People just know what to do with it. [9:58] Claude code is insanely powerful. [10:01] Go open up a terminal for the average [10:03] Fortune 500 employee and see how far [10:04] they get. [10:06] While it is powerful, [10:08] it does not offer that much affordance. [10:11] That's not a knock in on Anthropic, [10:13] but it is an opportunity for anybody who [10:15] wants to build on top. [10:17] And to create paths of least resistance [10:20] for your specific customers and their [10:22] specific problems so that it's just [10:25] brain dead simple for them to figure out [10:28] how to get to the outcome that they need [10:29] for their business. That's the concept [10:31] of affordance. [10:33] And then finally, the D in MAD is [10:36] diffusion. [10:37] And the diffusion gap is the opportunity [10:40] for companies building at the [10:41] application layer. The rate at which [10:44] capabilities are diffusing out into the [10:45] market is far shy of the rate at which [10:49] those capabilities are being created. [10:51] And every day that the foundation models [10:54] move faster than your average Fortune [10:56] 500 enterprise, that gap gets bigger [11:00] and that opportunity gets bigger. [11:03] So, for motes, [11:04] try to think customer back. [11:06] For affordance, try to think about [11:08] creating those paths of least resistance [11:10] for your customers. [11:12] And that diffusion gap, that represents [11:14] your opportunity. [11:16] Unless that slide from earlier with the [11:17] white space starting to fill up was [11:19] discouraging for anybody, may we remind [11:21] you [11:23] that no lead is safe. [11:25] There's [clears throat] this expression [11:25] in racing. You cannot pass 15 cars in [11:29] the sun, [11:30] but you can pass 15 cars in the rain. [11:33] And right now there is a torrential [11:35] downpour [11:36] of new capabilities coming out of the [11:38] foundation models. [11:40] Which means that no lead is safe, [11:42] but it also means [11:44] that anybody can win. [11:46] What a time to be alive. [11:48] And with that, I'll hand it off to [11:49] Sonya. [11:51] >> [applause] [11:56] [applause] [11:56] >> Thank you, Pat. And can I just say it's [11:58] so nice to see so many friendly faces in [12:00] the audience. There is an exceptional [12:02] group of people here today and I'm just [12:04] really happy to be part of this [12:06] ecosystem with all of you. And so, the [12:07] purpose of my section is to talk about [12:09] what's happening in AI right now, which [12:11] for 2026 is agents. [12:13] Okay, flashback to 2022. Show of hands, [12:16] does anybody here remember AutoGPT or [12:19] BabyAGI? Okay. Okay. So, these projects [12:22] were overnight hits on GitHub and what [12:24] they did was they took GPT-3, gave it [12:26] some tools, wrapped in a loop, and let [12:28] it run towards a goal. [12:30] And it was promising until you watched [12:32] those agents just fail over and over and [12:35] over again. [12:36] Kind of cute, kind of endearing, but [12:38] completely useless. [12:40] And I put this slide here to remind us [12:41] that you know, we all knew agents were [12:44] coming. Uh we could have seen it years [12:46] ago, [12:47] but back in 2022, the models just [12:49] weren't ready yet. [12:51] Fast forward to today, [12:53] something around the turn of the year [12:55] really changed. Suddenly we have agents [12:58] everywhere around us and they seem to [13:00] actually be working. [13:02] Two agents in particular have been home [13:04] runs. Claude code for the technical [13:06] crowd and open claw and all of its [13:08] lobster brethren, uh which democratized [13:10] agents to anybody with a phone. And so, [13:13] whether you are a hardcore engineer or a [13:15] normie, the punchline is that anybody [13:18] can create agents now. [13:20] And so, what we're seeing is people are [13:21] building agents for everything. There is [13:23] silly stuff like an open claw agent that [13:25] will literally snitch on your neighbors [13:27] for tax fraud. Please don't do this. Or [13:29] actually, maybe please do this. Uh [13:31] there's entrepreneurial stuff. Agents [13:33] running generative media campaigns to [13:35] sell construction services. [13:37] And then there's the professional layer. [13:39] I can tell you there's a huge race [13:40] internally at Sequoia for who can build [13:42] the best agents to do our jobs better. [13:45] So, what does it mean to be an agent? [13:48] Here is one possible definition. An [13:50] agent is a system that perceives its [13:52] environment, chooses actions, and [13:54] progresses autonomously towards a goal. [13:57] By the way, guys, I made this in C dance [13:58] by myself. I'm I'm very proud of it. [14:00] >> [laughter] [14:00] >> The video models have come a long way. [14:04] Uh and more specifically, I view agents [14:06] as having three functional components. [14:09] First is the ability to reason and plan. [14:12] Uh this is the baseline level of [14:14] intuition and the ability ability to [14:16] think on the fly. [14:18] Second is the ability to take actions. [14:21] This is tools, search, write, compile. [14:25] And then finally, the ability to iterate [14:27] towards a goal. This is the persistence [14:29] that gives agents the ability to [14:31] accomplish things over long time [14:33] horizons. [14:34] And so, agency combines these three [14:36] things. It is simply the ability to get [14:39] done. [14:42] If we boil the agents down into their [14:43] constituent components, the models, the [14:45] tools, the harnesses, each component has [14:48] progressed rapidly over the last year. [14:51] First, the models are the brain. This is [14:53] the most important thing that's [14:53] happened. The meter chart measures uh [14:56] how long a model can sustain progress on [14:58] a complex task without going off the [15:00] rails. And we've gone from the order of [15:03] tens of minutes a year ago to the order [15:05] of hours today. [15:07] And so, this is this is the most [15:08] important thing that's happened. The [15:09] models are finally getting capable [15:10] enough to sustain per- performance on [15:13] long horizon tasks. [15:17] Second, the tools are the arms and the [15:19] legs. [15:20] These give models access to things that [15:23] make us productive on the computer. The [15:24] terminal for file systems and dev tools, [15:27] iMessage, Slack, [15:29] uh web search, computer use, you name [15:32] it. [15:33] And the last two decades that we spent [15:34] building tools for humans have ended up [15:37] being uh able to port over to be [15:39] incredibly useful for for agents as [15:41] well. And there's a common refrain that [15:43] SaaS is dead. I think to the contrary, [15:45] the value of these tools is going to [15:47] explode as the number of agents using [15:49] them increases. [15:53] Models and tools give agents capability. [15:56] The harness is what gives them [15:57] persistence, the ability to stay on [15:59] task, adapt, and keep going. And that [16:02] feedback loop is now really starting to [16:04] crank. [16:05] Especially now with reinforcement [16:06] learning, we're giving these agents uh [16:08] we're taking them to driving school, [16:09] giving them uh [16:11] you know, training them uh in RL gyms, [16:14] and we're pushing performance in [16:15] different settings from mechanical [16:17] engineering to design to finance. [16:22] We're also seeing the early glimmers of [16:23] self-improvement or the machine building [16:25] the machine. [16:26] Uh for example, Andrej's other research [16:28] project uh improves research [16:30] autonomously towards a GPT-2 level model [16:33] in just 2 hours. [16:36] So, what does the world of agents [16:37] everywhere look like? [16:39] Agents exist on a sliding scale of [16:42] agenticness. Um and so, let's take [16:44] coding as an example. [16:46] In 2023, we had tab auto complete, one [16:49] AI assisting a human in line. This was [16:52] incrementally useful, fundamentally not [16:54] transformative. [16:57] We now have agentic development, one [16:59] human talking to an agent, instructing [17:01] it what to do, maybe managing a team of [17:03] agents. [17:04] But this paradigm is getting pushed [17:06] further. We're now seeing background [17:08] agents, async agents, agents spawning [17:10] sub agents. We think that async agents [17:13] in this whole paradigm is likely to [17:15] overtake the current paradigm in volume [17:17] just because of the amount of leverage [17:18] in the system. [17:20] And then finally, pushing the bleeding [17:22] edge of the frontier, what I call dark [17:23] factories, [17:25] um taking human review out of the system [17:27] completely. This sounds crazy, but I've [17:30] seen it happen in production including [17:32] with cybersecurity companies. It is [17:34] possible with good enough guardrails and [17:37] good enough engineering. [17:38] So, we're progressing up a scale of [17:40] agenticness. Agents are going from [17:42] little helpers that do a little amount [17:44] by your side to interns that need to be [17:47] managed to interns that manage [17:49] themselves. And eventually to interns [17:51] that can be trusted enough to to push to [17:53] prod without oversight. And so, that's [17:55] the evolution that's happening not just [17:57] in coding, but across all of agents. [18:01] Uh the most important takeaway uh for [18:03] the founders in this room is that [18:05] services is the new software. Pat's been [18:07] saying this for as long as I've known [18:08] him. Uh and our partner Julian, who's in [18:10] the audience today as well, uh published [18:12] a great article on this. And we've known [18:13] this for a long time, but I think it's [18:15] actually happening. So, in medicine, [18:16] you're able to hire an agent that [18:18] inspects your genome, gives you [18:20] personalized recommendations, can [18:21] prescribe you medication, uh recommend [18:23] you clinical trials. And law, you'll be [18:25] able to hire agents that can negotiate [18:27] contracts on your behalf, even perform [18:29] litigation and settle for you. [18:31] In math and the sciences, we're seeing [18:32] agents that can solve air those problems [18:34] or discover new superconductors. Like [18:36] how thrilling is that? [18:38] Or in uh in the consumer world, personal [18:40] agents that can manage your inbox for [18:43] you, your calendar, your finances, uh [18:45] file your taxes. [18:48] And we expect there's going to be agents [18:49] everywhere and that's in part because [18:51] hiring agents is so much easier than [18:53] hiring employees. [18:55] Humans are hard to scale. Infinites are [18:57] infinitely scalable with compute. [18:59] Humans are hard to keep happy, except [19:01] for me, I'm always happy. Uh agents are [19:04] low maintenance. [19:06] Uh humans are [19:08] uh humans are expensive. You pay them [19:09] salaries. [19:10] Uh you pay agents tokens. Generally, it [19:12] costs less to accomplish a task with [19:14] tokens than the equivalent in salary. [19:16] Today, humans are still generally [19:18] smarter, but the bitter lesson presses [19:19] on and soon agents will be smarter at [19:21] many things. And so, the point of this [19:23] slide is not that we humans are out of a [19:25] job. Uh I think a uniquely human trait [19:27] is adaptability. But we do expect the [19:30] deployment of agents across the [19:32] application layer to be swift and at an [19:34] unprecedented rate and scale because the [19:37] economics are so clear uh and because of [19:39] the inherent scalability of bits. [19:43] So, if you add all this up, the number [19:44] of agents is ballooning on some sort of [19:46] exponential, maybe super exponential. [19:49] And I think we're about to hit the point [19:50] where things get genuinely [19:51] [clears throat] strange. [19:52] Uh what happens when commerce happens [19:54] between agents? Can they pay each other? [19:57] What happens when agents can actually [19:58] negotiate the terms of a transaction [20:00] with each other? [20:01] Are we going to have swarms of agents [20:03] policing us, preventing things like [20:04] cybersecurity Armageddon? [20:06] All we know is the world is getting [20:08] weird extremely quickly. [20:11] And so I'll close by channeling my inner [20:13] my inner Bene Gesserit. [20:15] Uh long horizon agents are here. The [20:17] curves that they're on is very clear. [20:19] And for founders, I think everybody has [20:21] examples of people that are [20:23] accomplishing insanely hard timelines [20:25] thanks to AI. So Nathan from Zed [20:28] accomplished a three-year moonshot [20:29] project over the holidays by himself [20:31] with cloud code. Uh Bret Taylor rebuilt [20:34] Sierra over a weekend. Um [20:36] the Notion team rewrote 8 million lines [20:39] of code in just 6 weeks. And so [20:41] everybody has these examples of [20:42] compressed timelines. But I think very [20:44] few people outside of the AGI labs have [20:47] seen what happens when you take these [20:49] compressed timelines and you stack them [20:51] on top of each other. [20:52] And that's what's possible now. And so [20:55] whatever you could imagine building over [20:56] the next 100 years, we think is now [20:58] possible in 100 days thanks to agents. [21:01] I will pass it over to Konstantin. [21:03] Thank you, Sonia. [21:05] >> [applause] [21:07] >> All right. Thank you so much, Sonia, [21:09] Pat, for the brilliant overview, [21:11] analysis. In this section, we're going [21:13] to talk a little bit about what's next. [21:17] Uh so the goal here is we all know we're [21:19] in an AI age. [21:21] What's it going to look like? What's it [21:23] going to feel like? How's it [21:24] characterized? [21:26] Earlier in the presentation, Pat [21:29] bifurcated technological revolutions [21:32] between compute and communication. We're [21:34] going to do another bifurcation here for [21:35] types of work. [21:37] There is physical work. This is a [21:39] package on the Pony Express. This is a [21:41] satellite on a Falcon 9. [21:44] Work equals force times distance, [21:45] physical movement. And then there's [21:47] cognitive work. This is Pythagoras [21:49] coming up with his theorem. This is [21:51] DeepMind solving the pro- protein [21:54] folding problem. [21:55] Conscious thinking. [21:58] These are very different types of work. [22:01] But we believe that they're going to [22:02] follow a very similar pattern in [22:05] revolution. [22:07] So let's talk about physical work [22:08] because we've been through this [22:09] revolution with the Industrial [22:11] Revolution. [22:13] For the vast majority of human history, [22:17] all the work [22:19] or virtually all the work for serving [22:21] humans was done by some sort of muscle. [22:26] People or animals. People moving [22:28] something or an animal pulling the human [22:31] along. [22:33] This starts at 1700, but it goes back [22:35] millennia. [22:38] Then things started to change. [22:40] Water and wind. [22:42] Steam engines. [22:44] And then things accelerated. [22:47] Steam engines, combustion, electric [22:49] motors. [22:51] Today, 2026, [22:53] you could estimate that 99 plus percent [22:57] of all the physical work done on planet [23:00] Earth for humans is done by machine. [23:04] The plane that brought you here, [23:06] the manufacturing of all the goods in [23:07] this room, [23:09] all the transportation that sets up for [23:11] the pinnacle of the human experience [23:13] you're having right now. [23:15] Well, we think a similar pattern's going [23:18] to happen in cognition. [23:20] We're just a little earlier on. [23:23] >> [snorts] [23:23] >> So for most of human history, all the [23:25] thinking on planet Earth for humans was [23:28] done primarily by humans. Maybe a little [23:30] bit for animals. The the sheepdog [23:32] chasing the sheep. [23:34] Right? And there was this sliver on top [23:37] of mechanical work, the astrolabe or the [23:40] clock. [23:42] Now, over the past [23:45] couple hundred years, there was not a [23:46] lot of progress until [23:49] electronic computation. [23:52] And in the past 100 years, think about [23:54] all the trillions of calculations that [23:57] are happening at any given moment to [24:00] serve you the human. All of that work, [24:02] all of that cognitive work that's [24:04] happening to serve us at any given [24:06] moment. Trillions of calculations. [24:09] We believe [24:11] that the neural network [24:14] is the next big wave. [24:16] And that in the near future, 99.9% [24:20] of cognition on planet Earth will be [24:23] done by machines. [24:27] Well, [24:31] the parallel is pretty stark. [24:33] And the good news is we've been through [24:34] a revolution like this. [24:37] The cognitive revolution is going to be [24:38] a lot like the Industrial Revolution. [24:40] Just much, much bigger and much faster. [24:45] So what's it going to be like living in [24:47] this future? [24:48] I'd like to share some motivations for [24:51] this future in the form of four [snorts] [24:53] short stories. [24:56] The first story. [24:59] In the mid-1800s, America wanted to [25:01] build a grand monument to our first [25:04] president and our greatest war hero, [25:06] George Washington. [25:08] So we designed the tallest building in [25:10] the world at the time, the Washington [25:12] National Monument, [25:14] and we wanted to cap it with the most [25:17] precious metal in the world, 100 oz of [25:20] the most precious metal in the world. So [25:22] precious, in fact, that we put it on [25:24] display at Tiffany's in Manhattan. [25:27] That metal [25:29] was aluminum. [25:32] Within decades of the completion of the [25:35] Washington National Monument, [25:37] a young inventor came up with [25:39] electrolysis, [25:41] the process of separating aluminum from [25:43] dirt. [25:44] And within decades, [25:46] aluminum [25:48] was used to wrap our candies and our [25:50] sandwiches, [25:51] and then tossed into the trash. [25:55] Aluminum [25:57] is intelligence. [25:59] Electrolysis [26:01] is artificial intelligence. [26:04] We're about to enter a world where some [26:06] of the most precious skills [26:09] that took decades to earn, PhD-level [26:13] skills, [26:15] are so instantly invoked [26:18] that right after using them, you can [26:20] crumple them up and throw them right in [26:21] the trash. [26:26] Story number two. [26:28] We are entering a world of alien design. [26:32] The world as we see it today is all [26:34] about design for humans. [26:36] You know, it's it's been optimized in a [26:37] way that makes sense to our brains [26:39] because we are doing almost all the [26:41] cognition in the world. [26:44] Well, when machines do the cognition, [26:45] it's going to be a little different. In [26:47] 2006, NASA was optimizing an antenna for [26:51] a large space mission, satellite space [26:53] mission. [26:54] And traditionally, their antennas looked [26:56] like this. It was a beautiful geometric, [26:58] symmetrical pattern that optimized [27:01] surface area for some power constraints. [27:04] This time around, they said, "We're [27:05] going to hand it over to computer and [27:07] we're going to have an evolutionary [27:08] algorithm." A lot like reinforcement [27:10] learning. [27:11] The result, [27:13] this antenna right here. [27:16] Dramatically more productive. Not [27:19] intuitive to the human mind. [27:21] In this AI era, when we hand over [27:23] cognition to machines, we're going to [27:26] get results that are not intuitive to [27:27] us. [27:29] When AI's designing chips, cars, [27:32] buildings, [27:33] they might look dramatically different. [27:36] The world that we enter into, we have to [27:38] be open-minded because the AI is not [27:40] going to think like us. [27:43] It's going to have alien design. [27:46] The third motivation story is on [27:50] emerging sciences. Not emerging science. [27:53] We all know there's emerging science. [27:55] I'm talking about emerging sciences. [27:57] In the early Industrial Revolution, you [28:00] had great engineers like Newcomen and [28:03] Watt. And they perfected combustion [28:06] engines. [28:07] Basically, put a petrochemical into a [28:09] piston, ignite it on fire, millions, [28:12] billions of particles [28:13] explode, [28:15] move the piston, work. [28:17] For almost 100 years, all of that was [28:19] tinkering. [28:21] It was an engineer saying, "Ah, that [28:22] works a little bit better." Maybe [28:24] something you could see see like a [28:25] scaling law, but it was engineers [28:28] playing [28:29] with the product and seeing how they [28:31] could improve it a little bit. [28:34] Over 120 years after [28:37] Sadi Carnot came around [28:39] and formalized this in a new science, [28:42] thermodynamics. [28:44] He said, "Wait a second. [28:46] There are millions or billions of [28:47] particles. We can actually formalize [28:49] what that all looks like." [28:50] In this case, [28:52] there are billions of neurons, trillions [28:54] of tokens. [28:55] Right now, we're in the tinkering phase [28:57] of AI. Even if we think it's an [28:58] understood science, it's not. [29:00] In the future, we will have a science as [29:03] fundamental as thermodynamics introduced [29:06] in the next couple decades. Someone in [29:07] this room might come up with that [29:08] science. [29:10] And that science will be taught in high [29:12] schools. It will be that fundamental. [29:15] And it will help us master AI. It will [29:17] even help us master consciousness. [29:21] Fourth story. [29:24] The art of unreason. [29:27] So for the vast majority of human [29:29] history, [29:30] tens of thousands of years, art has been [29:32] a progression towards realism. [29:36] This is a cave painting from about [29:38] 25,000 years ago. [29:41] Egyptian hieroglyphs. [29:44] Greek [29:45] pottery. [29:47] Uh Renaissance paintings. [29:50] uh [29:51] a grand transformation toward realistic [29:54] art. Just look at the difference [29:56] over tens of thousands of years, the [29:59] triumph of humanity. [30:01] And then, [30:03] engineering came along. [30:06] The daguerreotype, early photography. [30:10] And all of a sudden, what was spent [30:11] decades of life to perfect the skill of [30:15] getting every brushstroke perfect, [30:17] done. [30:19] So, how did the world react? They [30:20] thought that painting was over. [30:23] Oh, that's it. The machine can do it [30:24] better than any human. [30:26] Art is Art is ended. [30:30] Well, what happened? How did humans [30:32] respond? [30:34] Humans responded [30:35] by saying, [30:37] "Was the purpose of this art to capture [30:39] the moment in the way the eye sees it? [30:43] Or was it to capture the moment in the [30:45] way the heart and the soul sees it?" [30:49] Impressionism, expressionism, [snorts] [30:52] cubism, [30:54] neo-expressionism. [30:56] All these new forms of art are how [30:58] humanity responded to this dramatic [31:01] change in science. [31:05] 2,500 years ago, [31:08] Greek philosopher Protagoras [31:11] wrote, [31:12] "Man is the measure of all things." [31:17] What he meant [31:18] is that nothing [31:20] in a vacuum [31:22] has value to humans. [31:25] Not aluminum, [31:26] not art, [31:28] not intelligence. [31:30] >> [snorts] [31:30] >> It only has value [31:33] because of the experience. [31:37] AI can do the work. AI will do the work. [31:41] But, [31:42] only the human connection [31:44] can give you a reason [31:46] to care. [31:50] That's why we're all in this room today. [31:52] A decade from now, work is going to be [31:54] dramatically different. Things are going [31:56] to change so much, but the one thing [31:58] that will be constant [32:00] is the relationships that you form today [32:02] with the person right next to you will [32:04] endure. [32:06] That's what you're going to look back [32:07] on. That's what's going to be valuable [32:08] from today. [32:09] So, I encourage you to form those [32:11] relationships with the people next to [32:13] you. [32:14] Enjoy your time together at this AI [32:16] Ascent, [32:17] and really lean into what makes us most [32:19] human. [32:22] >> [applause]