年费999美元文章:AI代理的未来影响行业及公司

Agentic Utilities代理实用程序

Always On, Never Prompted始终在线，从不提示

Citrini西特里尼

Mar 25, 20262026年3月25日

∙ Paid已付费Introduction介绍

Here at Citrini Research, we’ve been talking about “Agents” for as long as we’ve been writing about AI.在 Citrini Research，自从我们开始撰写有关人工智能的文章以来，我们就一直在谈论“代理”。

Back in May 2023, in our Artificial Intelligence: Global Equity Beneficiaries piece, we predicted that AI adoption (and its investing implications) would follow a three-phase process:早在 2023 年 5 月，我们在 《人工智能：全球股票受益者 》一文中预测，人工智能的采用（及其对投资的影响）将遵循三个阶段的过程：

Phase I: Global Data Center Hyperscaling (“infrastructure”)第一阶段： 全球数据中心超大规模（“基础设施”）
Phase II: The Democratization of AI/ML (“commodification”)第二阶段： 人工智能/机器学习的民主化（“商品化”）
Phase III: Integration & Specialization (you’re here now)第三阶段： 整合与专业化 （您现在所在位置）

Here’s how we described the vision at the time:以下是我们当时对这一愿景的描述：

he third phase is marked by the deeper integration of AI and ML into various industries and aspects of society, as well as further specialization within the field of AI itself. The technologies described as “artificial intelligence” may not resemble the LLMs we currently use, and one or more services will provide “agents” capable of carrying out tasks without being prompted directly (an “AI assistant”).”第三阶段的标志是人工智能和机器学习更深入地融入各个行业和社会各个方面，以及人工智能领域本身的进一步专业化。被描述为“人工智能”的技术可能与我们目前使用的机器学习方法（LLM）有所不同，并且一种或多种服务将提供无需直接指令即可执行任务的“代理”（即“人工智能助手”）。

This roadmap has proven prescient, though perhaps a better name for Phase III would be the “Agentic Era”.事实证明，这份路线图极具前瞻性，不过或许将第三阶段命名为“ 代理时代”会更好。

We start today’s piece with this callback for a reason (not just to pat ourselves on the back). The reason is that we, like many others, have envisioned the coming of the agents pretty much straight out of the gate; AI was never going to be constrained to a chat bot.我们今天以这段回顾作为文章的开头是有原因的（并非只是为了自我吹嘘）。原因是，和许多人一样，我们从一开始就预见到了智能体的出现；人工智能的应用范围绝不会局限于聊天机器人。

In our follow-up piece, Less Deus More Machina, we zeroed in on “AI Losers” – companies and industries most vulnerable to disruption with the proliferation of AI (and, importantly, agentic AI). Our “standalone short” basket of 26 losers worked fairly well against a rising market for the following two years, albeit with a good deal of choppiness along the way. But, the waterfall really began late last year – just as the “agents” arrived.在我们的后续文章《 少些神明，多些机械》 中，我们重点关注了“人工智能输家”——那些最容易受到人工智能（尤其是智能体人工智能 ）普及冲击的公司和行业。在接下来的两年里，我们构建的包含 26 只输家的 “独立做空”投资组合在市场上涨的情况下表现相当不错，尽管期间波动较大。但真正的崩盘始于去年年底——就在“智能体”出现之时。

We’ve reached an inflection. AI Agents have gone from an abstract concept to an actual service with expansive real world capabilities that can be directed from a simple iMessage. For the “losers”, the theoretical risks that were enough to stoke fear for investors have rapidly become tangible risks that fundamentally alter management outlooks.我们已经到了一个转折点。人工智能代理已经从抽象概念发展成为一项实际服务，拥有广泛的现实世界功能，并且可以通过简单的 iMessage 进行操控。对于那些“失败者”来说，曾经足以引发投资者恐慌的理论风险，如今已迅速转化为切实存在的风险，从根本上改变了管理层的预期。

The concept of an “AI Agent” went from buzzword to reality in November 2025 – reaching escape velocity with the release of ClawdBot (now OpenClaw, thanks to Anthropic’s lawyers). Until recently, the coding tools suffered the same flaw as every other hacked-together “AI” SaaS app: the call and response model. Due to degrading context windows, models couldn’t store history of past events so there was only so much that a single run could do. You were still limited by the command line, but not anymore.“人工智能代理”的概念在 2025 年 11 月从流行语变为现实——随着 ClawdBot（现在更名为 OpenClaw，这要感谢 Anthropic 公司的律师）的发布，它的发展速度达到了前所未有的高度。直到最近，这些编码工具都和其他所有拼凑起来的“人工智能”SaaS 应用一样，存在着同样的缺陷：调用-响应模型。由于上下文窗口的退化，模型无法存储过往事件的历史，因此单次运行所能完成的任务非常有限。虽然你仍然会受到命令行的限制，但这种情况已经不复存在了。

In January, OpenClaw – an open-sourced, self-hosted agentic library – became the poster child for agentic technology. As shown by GitHub’s “Star History”, the adoption has been astronomical in just two months.今年一月，开源的、可自行托管的智能体库 OpenClaw 成为了智能体技术的典范。正如 GitHub 的“Star History”所示，短短两个月内，它的用户数量就呈爆炸式增长。

Along with exploding Github references, OpenRouter data shows just how dramatically OpenClaw has accelerated token usage within its product category.除了 GitHub 引用量激增之外，OpenRouter 的数据也显示，OpenClaw 在其产品类别中极大地加速了代币的使用。

OpenClaw went viral because it had the bright idea of storing context in a big text file and looping through new AI calls. Early users were running their agents overnight, and a new ecosystem of skills and plugins boosted its network effects. And, for just the cost of a Mac Mini, you could run it all via iMessage. We must reiterate that this is the core reason behind the ClawdBot/OpenClaw Mac Mini craze: it has nothing to do with the hardware. Most of these functions are run in the cloud rather than locally, although longtime readers know our thoughts on where that’s headed.OpenClaw 之所以迅速走红，是因为它巧妙地将上下文信息存储在一个大型文本文件中，并循环执行新的 AI 调用。早期用户甚至在夜间运行他们的智能体，而全新的技能和插件生态系统也进一步提升了它的网络影响力。更令人惊喜的是，只需一台 Mac Mini 的价格，你就能通过 iMessage 运行所有程序。我们必须重申，这才是 ClawdBot/OpenClaw 在 Mac Mini 上风靡的核心原因：它与硬件本身无关。虽然大部分功能都在云端运行，而非本地，但我们的老读者都知道我们对此的看法。

In this new paradigm, Jensen Huang is floating the idea of providing “token based compensation” to attract engineering talent. Jensen also recently commented that he would be ‘deeply alarmed’ if one of his engineers earning $500,000 a year did not consume at least $250,000 in tokens. While we’re cognizant of the fact that this is like a baker saying you should be eating a dozen donuts a day, there’s a lot of truth in it as well.在这种新模式下，黄仁勋提出了“代币化薪酬”的概念，旨在吸引工程人才。黄仁勋最近还表示，如果他手下年薪 50 万美元的工程师没有消费至少价值 25 万美元的代币，他会“深感担忧”。虽然我们知道这听起来像是面包师建议你每天吃一打甜甜圈，但其中也蕴含着不少道理。

Homebrew Agents are still in “move fast and break things” mode. Agents are vibecoding, answering emails, dumping crypto wallets, and folding proteins. And companies have taken notice…自研智能体仍然处于“快速行动，打破常规”的模式。它们能够感知周围环境、回复邮件、转移加密钱包，甚至折叠蛋白质。而各大公司也已经注意到了这一点……

First, “Agentic AI” is now widely acknowledged as a critical market narrative.首先，“智能体人工智能”现在已被广泛认为是重要的市场叙事。

Second, picking winners and losers in this space will be a whole lot harder than simply seeing who is talking about it on earnings calls.其次，在这个领域挑选赢家和输家，要比仅仅看看财报电话会议上谁在谈论它要困难得多。

Nevertheless, we expect this to be a powerful narrative for true “agentic winners” – particularly if it reshapes the investment case for companies in the dumps. As always, we love a great bottom.然而，我们预计这将为真正的“代理赢家”带来强有力的启示——尤其是在它重塑那些处于低谷公司的投资价值时。一如既往， 我们热爱触底反弹 。

We have been slowly adding to our allocation names that are aligned with the “Agentic Utility” layer – we are currently long AKAM, FSLY, CRCL and NET in the Citrindex. We’ve bought all of these when they were widely viewed as being entirely separate from the AI trade. Now, we think the moves in these names speak to how quickly stories and sentiment can reverse when the market recognizes an AI angle.我们一直在逐步增加与“智能代理实用工具”层相关的股票配置——目前我们在 Citrindex 指数中持有 AKAM、FSLY、CRCL 和 NET 的多头头寸。我们买入这些股票时，它们普遍被认为与人工智能交易完全无关。现在，我们认为这些股票的走势表明，当市场意识到人工智能的利基时，市场情绪和看法可能会迅速逆转。

To understand where other agentic winners may lie, let’s first establish a framework for where to look for any novel beneficiaries of the new era of supercharged token consumption.为了了解其他潜在的受益者可能在哪里，我们首先要建立一个框架，以便寻找新时代超高速代币消费的任何新受益者。

Framework: Agentic Utilities框架：代理实用程序

Let’s start with a metaphor. In the chatbot era of AI, ChatGPT could help you plan dinner. It could help you pick out a menu, find recipes, and create a shopping list. It might even serve you ads that help you buy the ingredients. All that may be super helpful, but at the end of the day, it couldn’t actually make you dinner. Ultimately, you must use your own agency to drive to the store, slice the onions, and plate the chicken parm.我们先举个例子。在人工智能聊天机器人时代，ChatGPT 可以帮你计划晚餐。它可以帮你挑选菜单、查找食谱、创建购物清单，甚至还能推送一些帮你购买食材的广告。所有这些都非常有用，但归根结底，它并不能真的帮你做好晚餐。最终，你还是得自己开车去超市、切洋葱、摆盘鸡肉帕尔玛。

But imagine a world in which it can. Imagine if you had a service at your fingertips that could hop in a car, buy your groceries, pick up your laundry, and stop at the bank on the way home. It wouldn’t just stop at one grocery store, it would send ten cars out to every single grocery store in town to make sure you’re getting your Cheerios at the lowest price. It’s not free (it still chews tokens – and the “juice” of price comparison must be worth the compute), but as far as services go, it’s pretty darn useful – and in fact, you might use it a lot.但想象一下，如果这一切真的发生了。想象一下，如果你的指尖轻触就能拥有这样一项服务：它可以开车帮你买菜、取衣服，回家路上还能顺便去趟银行。它可不只是去一家超市，而是会派出十辆车跑遍镇上每一家超市，确保你买到的麦片价格最低。这项服务并非免费 （它仍然需要消耗代币——而且价格比较带来的“收益”肯定也需要计算），但就服务而言，它确实非常实用——事实上，你可能会经常用到它。

But what does this world actually look like?但这个世界的真实面貌究竟是怎样的呢？

Second, the businesses that cater and optimize for these new task rabbits will see an influx in volume. Meanwhile, those that cater solely to human touch will lose market share.其次，那些迎合并优化这些新型“任务型员工”需求的企业将会迎来业务量的激增。与此同时，那些只注重人际互动的企业将会失去市场份额。

Finally, these swarms of agents become a risk to individuals and businesses alike. Sheer overload might resemble a stampede even if intentions are good – even worse, they are just as easily directed towards nefarious ends.最终，这些数量庞大的代理会对个人和企业都构成威胁 。即使出发点是好的，数量过多也可能引发类似踩踏事件的混乱——更糟糕的是，它们也很容易被用于不法目的。

For the immediate future, this metaphor applies mostly to the digital realm. The physical embodiment of AI via autonomous driving and robotics may be closer than we think, but we’ll hold off on this for now. The entire digital landscape of infrastructure, commercial interface, and security must adapt to a new paradigm.就目前而言，这种比喻主要适用于数字领域 。人工智能通过自动驾驶和机器人技术实现物理化或许比我们想象的更近，但我们暂且搁置这一话题 。整个数字基础设施、商业界面和安全领域都必须适应新的范式。

We classify the winners into three categories.我们将获奖者分为三类。

1) Infrastructure: The Agentic Internet has a different structure than the one we are currently used to, both in terms of pathing and bandwidth. Digital plumbing needs to accommodate a boom in agentic traffic. This has unearthed legacy names that are already seeing inflections in both revenue and guidance that prove the demand has already arrived, and is rapidly growing.1）基础设施：代理互联网的结构与我们目前习惯的互联网截然不同，无论从路径还是带宽方面来看都是如此。数字基础设施需要适应代理流量的激增。这使得一些传统企业重新焕发活力，它们的营收和业绩预期都出现了显著的转变，这证明市场需求已经到来，并且正在快速增长。

2) Ecosystem: Simultaneously, a new customer category has emerged. Instead of selling to consumers (B2C) or businesses (B2B), companies will offer Agent (B2A) services.2）生态系统：与此同时，一种新的客户群体出现了。企业不再直接面向消费者（B2C）或企业（B2B）销售产品，而是提供代理（B2A）服务。

Agentic Utilities are the first vendors in the B2A vertical – these are companies offering infrastructure and services to be utilized by agents that can’t be “vibecoded” away – for regulatory, operational, or path-dependent reasons.Agentic Utilities 是 B2A 垂直领域的首批供应商—— 这些公司提供基础设施和服务，供代理商使用，由于监管、运营或路径依赖的原因，代理商无法通过“vibecoded”方式摆脱。

Agents need ways to interact with the real world, and that includes payment rails. Stablecoins and agentic wallets are being quickly adopted, with non-humans making financial transactions using both crypto and the SWIFT system. Even robots can’t escape the dollar’s reserve currency status.智能体需要与现实世界互动的方式，这其中就包括支付渠道。稳定币和智能体钱包正在迅速普及，非人类实体也开始使用加密货币和 SWIFT 系统进行金融交易。 即使是机器人也无法摆脱美元的储备货币地位。

The evolution of the ecosystem will wind through the buildout of agentic plumbing (we’re seeing that now with the moves in the CDNs) and then begin pricing in their implementation. Value will accrue first to the picks and shovels, and then to the APIs in demand, the agentic harnesses (i.e. where the attention is) and the agents themselves.生态系统的演进将经历代理基础设施的构建（我们目前已从 CDN 的举措中看到这一点），然后开始对其实现进行定价。价值将首先体现在基础架构上，然后是需求旺盛的 API、代理工具（即用户关注的焦点所在）以及代理本身。

3) Governance: And, the Agentic Internet needs new protections. Current AI capabilities make it possible for a rogue agent to break the internal systems at Meta, and that is the dumbest that agent will ever be. Every single company is asking “where is the killswitch, and how do we know when to use it?” This puts observability and counter-AI solutions in play. These names were thrown out with the bathwater when the entire software industry developed a new risk premium in under a quarter.3）治理：智能体互联网需要新的保护措施。目前的人工智能能力使得一个失控的智能体能够攻破 Meta 的内部系统，而这已经是该智能体所能达到的最愚蠢的程度了。每家公司都在问：“ 终止开关在哪里？我们如何知道何时使用它？” 这就使得可观测性和反人工智能解决方案变得至关重要。然而，当整个软件行业在不到一个季度的时间里就发展出新的风险溢价时，这些概念却被一带而过。

The companies that address these three problems today are laying the groundwork for the agentic future – we call them Agentic Utilities.如今致力于解决这三个问题的公司正在为智能体的未来奠定基础——我们称它们为智能体公用事业公司。

1) Infrastructure1）基础设施

We’ve escaped the command line interface and it’s going to change the structure of the internet.我们已经摆脱了命令行界面，这将改变互联网的结构。

Cisco (CSCO US) estimates Agentic AI generates up to 25x more network traffic than a chatbot. That figure compounds as more robust agents run in an “almost always on” fashion. Or in the words of A10 Networks (ATEN US), “AI is a traffic problem before it’s a compute problem”.思科（CSCO US） 估计，智能代理 AI 产生的网络流量是聊天机器人的 25 倍。随着功能更强大的代理以“几乎始终在线”的方式运行，这一数字还会进一步增加。或者用 A10 Networks（ATEN US） 的话来说，“AI 首先是一个流量问题，然后才是一个计算问题”。

Training and inference determine what a model knows. Traffic determines whether it’s useful at scale. Once AI starts doing real work, every answer becomes a routing problem.训练和推理决定了模型掌握的知识。流量决定了模型能否大规模应用。一旦人工智能开始执行实际任务，每个答案都变成了一个路径规划问题。

This traffic growth is not evenly distributed. Demand is shifting from centralized hyperscaler training complexes towards edge compute for real-time inference. As more agentic work happens behind the scenes without any user interaction, the pathing of internet traffic is fundamentally changing.这种流量增长并非均匀分布。需求正从集中式超大规模数据中心的训练集群转向用于实时推理的边缘计算。随着更多无需用户交互的后台智能计算工作发生，互联网流量的路径正在发生根本性的变化。

The City of Austin shows exactly how this will work.奥斯汀市政府具体展示了这种模式是如何运作的。

Traffic in Austin is terrible, and it’s baked into the structure of the city. There are two major “north-south” roads that run through town: Mopac and Interstate 35.奥斯汀的交通状况糟糕透了，这几乎是这座城市结构固有的问题。有两条主要的南北向道路贯穿市区：莫帕克高速公路和35号州际公路。

The infrastructure failure is that there are no major “east-west” roads that go through the city, which turns the road system into a hellscape of flyways and on-ramps. This causes major congestion on both sides of the city as there’s no easy way to cross. The same bottlenecks will appear in our digital infrastructure and will demand new solutions.基础设施的缺陷在于缺乏贯穿城市的主要东西向道路，导致道路系统变成了由高架桥和匝道组成的混乱地带。由于缺乏便捷的过境方式，城市两侧都出现了严重的交通拥堵。同样的瓶颈也将出现在我们的数字基础设施中，需要新的解决方案。

Orthogonal OSI (Open Systems Interconnection)正交 OSI（开放系统互连）

The traditional internet was designed with human-machine interaction in mind. This is north-south traffic: you type in a URL, your browser sends a request to a server, and then the server responds with content.传统互联网的设计以人机交互为核心。这是南北向的通信模式：你输入网址，浏览器向服务器发送请求，然后服务器返回内容。

The agentic shift causes a different pathing of internet traffic. When AI agents are given a task, they then operate “under the surface” through API access, negotiating with other agents, while maintaining persistent sessions.代理的转变导致了互联网流量路径的改变。当人工智能代理被赋予任务时，它们会通过 API 访问在“底层”运行，与其他代理进行协商，同时保持持久会话。

Agentic traffic will have more “east-west” movement, where 80% of traffic will move laterally between servers, GPUs, and data centers. As inference scales across multiple data centers, the demand for “inter-server” traffic and corporate networks will multiply.代理流量将更多地呈现“东西向”流动，其中 80% 的流量将在服务器、GPU 和数据中心之间横向移动。随着推理规模扩展到多个数据中心，对“服务器间”流量和企业网络的需求将成倍增长。

“The global agentic-AI M2M [machine to machine] traffic is expected to increase from 66 EB/month in 2025 to 537 EB/month in 2034”“预计全球基于人工智能的机器对机器（M2M）通信流量将从 2025 年的每月 66 EB 增长到 2034 年的每月 537 EB。”

– Nokia Global Traffic Report– 诺基亚全球流量报告

The types of packets aren’t going to be the same. Human-machine interactions tend to be short sessions that open and close. Agentic traffic demands a different kind of internet: always-on connections, higher bitrates, and continuous access to inference clusters.数据包的类型将截然不同。人机交互通常是短暂的会话，开启和关闭即告结束。而智能体流量则需要一种不同的互联网：始终在线的连接、更高的比特率以及对推理集群的持续访问。

The demand for east-west traffic is creating a structural change to the internet’s plumbing, and a structural demand shift for managing that bandwidth. This shift should continue to benefit the physical players that make up the backbone of the internet.东西向流量的需求正在对互联网的底层架构造成结构性改变，并对带宽管理提出结构性需求转变。这种转变应会持续惠及构成互联网骨干网的物理运营商。

Cisco is involved in multiple layers of the OSI model, with leadership focused on the emergent “always on” model of the agentic internet. They are responding to a fresh cycle of enterprise networking upgrades just as investor enthusiasm is waning on whether AI networking tailwinds have peaked.思科参与了 OSI 模型的多个层面，其领导层专注于新兴的 “始终在线” 的智能互联网模式。他们正积极应对新一轮的企业网络升级浪潮，而此时投资者对人工智能网络发展势头是否已经达到顶峰的热情却正在消退。

Cisco has developed both the SiliconOne G300 – a 102.4 Tbps switch and an AgenticOps platform. They’re leveraging their acquisition of Splunk with a new observability offering while building a security wrapper for the new agentic attack surface.思科开发了 SiliconOne G300（ 一款 102.4 Tbps 交换机）和 AgenticOps 平台。他们正利用对 Splunk 的收购，推出新的可观测性产品，同时为新的代理攻击面构建安全防护层。

Ciena (CIEN US) is another standout in this area, as it is the optical backbone of the east-west traffic expansion.Ciena（CIEN US）是该领域的另一家杰出企业，因为它是东西向流量扩展的光纤骨干网。

Ciena recently increased capex by about 2.5x the historical average to build out capacity; the company explicitly states that it’s meeting demand rather than anticipating it.Ciena 最近将资本支出增加了约 2.5 倍，以扩大产能；该公司明确表示，它正在满足需求， 而不是预测需求。

There’s also an agentic platform kicker – the company’s Blue Planet division has built out an Agentic AI framework for network operations, which has already been adopted by Lumen as of October 2025.此外，还有一个智能平台方面的亮点——该公司旗下的 Blue Planet 部门已经构建了一个用于网络运营的智能 AI 框架，该框架已于 2025 年 10 月被 Lumen 采用。

Hewlett Packard Enterprise (HPE US) also has clear inroads into the AI network infrastructure trade through its recent acquisition of Juniper Networks.惠普企业（HPE US）最近收购了瞻博网络 ，也明显进军人工智能网络基础设施领域。

HPE’s networking segment grew 152% YoY, boosted by the Juniper acquisition, and is now sitting at 30% of total revenues and half of operating profits.受 Juniper 收购的推动， HPE 的网络业务同比增长 152%， 目前占总收入的 30% 和营业利润的一半。

That’s not even touching their existing “on-prem” business line. HPE Private Cloud AI offers a full stack subscription (down to the chips) so that clients can run inference without exposing proprietary data to any public networks.这甚至还没涉及到他们现有的“本地部署”业务线。HPE 私有云 AI 提供全栈订阅服务（涵盖芯片层面），让客户无需将专有数据暴露给任何公共网络即可运行推理。

The Street Signs and the Roadmap街道标志和路线图

To extend the Austin metaphor (sorry), you can build all the east-west highways you want…but they won’t accomplish their goal if nobody driving on them can read the street signs or look up an address. Before a single packet of agentic traffic traverses CIEN’s optical backbone or reaches an inference cloud the agent needs to resolve a domain name and the connection needs a routable IP address.借用奥斯汀的例子（抱歉），你可以修建所有东西向高速公路……但如果路上的司机无法看懂路标或查找地址，这些高速公路就无法实现其目的。在任何代理流量数据包通过 CIEN 的光骨干网或到达推理云之前，代理需要解析域名，并且连接需要一个可路由的 IP 地址。

This brings us to an interesting angle – two distinct chokepoints sitting on legacy internet infrastructure names.这就引出了一个有趣的角度——两个不同的瓶颈位于传统的互联网基础设施名称上。

A10 Networks (ATEN US) started out providing Application Delivery Controllers (ADCs), balancing loads between servers and users. These are rapidly being displaced by virtualization and programmable silicon.A10 Networks（ATEN US）最初提供应用交付控制器（ADC），用于平衡服务器和用户之间的负载。但这些产品正迅速被虚拟化和可编程芯片所取代。

Their upside is embedded in the carrier grade network management infrastructure that sits directly in the path of key agentic demand: unique sessions for agents and on-device inference. Every agent that wants to maintain a persistent connection with another server, call an API or coordinate with a physical device in the real world requires an addressable endpoint. The problem is there are only so many IP addresses to go around – roughly 4.3 billion. We actually ran out years ago.它们的优势在于运营商级网络管理基础设施 ，该基础设施直接服务于关键的代理需求：代理的独立会话和设备端推理。每个想要与另一台服务器保持持久连接、调用 API 或与现实世界中的物理设备协调的代理都需要一个可寻址的端点。问题在于 IP 地址的数量是有限的——大约只有 43 亿个。实际上，我们几年前就已经用完了。

The workaround is NAT (Network Address Translation), which lets thousands of devices share a single public IP address by managing sessions behind the scenes. A10 handles the NAT table at the scale of entire telecom networks, and they’re simultaneously building out IPv6 transition infrastructure for the day (still years away) when the internet moves to a larger address space. A10’s Thunder CGN helps manage the (now scarce) IPv4 addresses while building out the scale for IPv6. The company is focused on growing the cybersecurity wrapper around their hardware and projects 65% of revenue will come from this domain.这种变通方案是使用 NAT（网络地址转换），它通过在后台管理会话，允许数千台设备共享同一个公网 IP 地址。A10 公司负责处理整个电信网络规模的 NAT 表，同时也在构建 IPv6 过渡基础设施，以应对互联网迁移到更大地址空间的那一天（虽然还有数年时间）。A10 的 Thunder CGN 解决方案有助于管理（如今已十分稀缺的）IPv4 地址，同时为 IPv6 的扩展做好准备。该公司专注于发展其硬件的网络安全保护方案，并预计 65% 的收入将来自该领域。

While A10 manages scarce IPv4 addresses and builds out IPv6 scale for the explosion of agentic endpoints, VeriSign (VRSN US) is the other side of that same coin. If A10 manages the street signs, VRSN manages the roadmap.A10 负责管理稀缺的 IPv4 地址，并为代理终端的爆炸式增长构建 IPv6 规模；而 VeriSign（VRSN US） 则是同一枚硬币的两面。如果说 A10 管理的是路标，那么 VRSN 则管理的是路线图。

VeriSign operates the authoritative registry for .com and .net and processes 450 billion DNS queries a day. Every web request an agent makes starts with DNS resolution. CEO James Bidzos has explicitly flagged agentic AI as a driver of DNS utilization on the last two earnings calls, and the company is teasing new security services for the first time in a decade.VeriSign 运营着 .com 和 .net 的权威域名注册管理机构，每天处理 4500 亿次 DNS 查询。用户发出的每个网络请求都始于 DNS 解析。首席执行官 James Bidzos 在最近两次财报电话会议上都明确指出，智能代理 AI 是 DNS 使用率增长的主要驱动因素，而且该公司十年来首次预告将推出新的安全服务。

The traditional internet required one DNS lookup per page load. Agents becoming more commonplace also means their workflows fan out across ten APIs, generating ten lookups before it’s even started reasoning. Multiply that by the always-on, persistent session model described throughout this piece and you can see the bull case for VRSN – a structural increase in DNS query volume compounding with every new agent deployed.传统互联网每次页面加载都需要一次 DNS 查询。随着代理的日益普及，它们的工作流程会扩展到十几个 API，在开始推理之前就会产生十次查询。再加上本文中描述的始终在线的持久会话模型，就不难理解 VRSN 的优势所在——DNS 查询量会随着每个新代理的部署而结构性地增加。

VeriSign has positioned domain names as globally unique, stable, human-readable identifiers for verifying digital content, which they argue is critical for agentic AI and valuable for combating misinformation and deepfakes. VRSN is an amazing business model – 67% operating margins with virtually zero customer acquisition costs. Now, imagine if the business gets priced for growth again (it’s not that hard, VRSN went from about six dollars to $240 during the dot com bubble).VeriSign 将域名定位为全球唯一、稳定且易于理解的标识符，用于验证数字内容。他们认为这对于智能人工智能至关重要，并且对于打击虚假信息和深度伪造技术也极具价值。VRSN 的商业模式非常出色——运营利润率高达 67%，而获客成本几乎为零。现在，想象一下，如果这家公司的股价再次反映出增长潜力（这并不难，在互联网泡沫时期，VRSN 的股价就从大约 6 美元飙升至 240 美元）。

Location, Location, Location位置，位置，位置

Agentic AI priorities are shifting from raw compute to speed. Agents don’t have the patience of a human waiting for a web page to load – workflows require quick answers to maintain a constant reasoning loop, whether that loop is to review code or ensure two robots don’t collide.智能体人工智能的关注点正从原始计算能力转向速度。智能体不像人类那样有耐心等待网页加载——工作流程需要快速响应以维持持续的推理循环，无论这个循环是用于代码审查还是确保两个机器人不会相撞。

What about content that hasn’t yet been generated? What if you need a conversational voice interface? What if you need to coordinate a fleet of robots?如果内容尚未生成怎么办？如果需要对话式语音界面怎么办？需要协调一支机器人队伍怎么办？

You move compute closer to the endpoint.你将计算位置移近终点。

The AI infrastructure narrative has been primarily focused on centralized compute hubs, where hyperscalers are building massive data centers close to cheap energy inputs. The other extreme and somewhat speculative end of the spectrum is inference-on-device, where AI models can run directly on local devices.人工智能基础设施的讨论主要集中在集中式计算中心，超大规模数据中心运营商在靠近廉价能源供应地的地方建设大型数据中心 。而另一种极端且带有一定推测性的方案是设备端推理，即人工智能模型可以直接在本地设备上运行。

(As we’ve long argued, we do see an eventuality in which meaningful compute can be held in your own hands. Developing technologies like that of Taalas (private) is taking model weights and placing them in the silicon alongside on-chip memory, and could accelerate this future.)（正如我们长期以来所论证的，我们确实预见到未来有一天，真正有意义的计算能力将掌握在个人手中。像 Taalas（私有公司）这样的技术正在开发中，它将模型权重与片上存储器一起存储在硅芯片中，这可能会加速这一未来的到来。）

But between “big ass data center” and “pocket AGI” lies a middle ground of pre-existing infrastructure that has been retooled for the agentic revolution. And, with the AI narrative focused on the ends of the barbell, these names haven’t yet seen a large markup in response to agentic traffic demands.但介于 “巨型数据中心” 和 “掌上通用人工智能” 之间 ，存在着一个中间地带，即对现有基础设施进行改造以适应智能体革命。 而且，由于人工智能的叙事主要集中在两端，这些名称尚未因智能体流量需求而出现大幅上涨。

CDNs: “Compute” Delivery NetworksCDN：计算分发网络

When the World Wide Web hit the mainstream in the mid 90’s, it revealed the flaw in a centralized internet. If a website’s images were hosted on a single server 2000 miles away, then the end user would have to wait entire seconds for the page to load.20 世纪 90 年代中期，万维网开始普及，但也暴露了中心化互联网的缺陷。如果一个网站的图片托管在 2000 英里之外的单个服务器上，那么最终用户就需要等待几秒钟才能加载页面。

There was the “hug of death” if your site went viral. Thousands of users would flood a website from a Slashdot link, causing the server to overload, then crash. It only got worse – when the Ken Starr Report was released in 1998, major media and government websites went down for the rest of the business day.如果你的网站爆红，就会遭遇“死亡之吻”。成千上万的用户会通过 Slashdot 上的链接涌入网站，导致服务器过载崩溃。情况只会更糟——1998 年肯·斯塔尔报告发布后，各大媒体和政府网站当天全部瘫痪。

Content delivery networks (CDNs) were created to help decentralize the internet. Website owners could “cache” images on servers close to the end-user’s location, allowing faster load times and steadier connectivity.内容分发网络（CDN）的创建是为了帮助实现互联网的去中心化。网站所有者可以将图像“缓存”在靠近最终用户位置的服务器上，从而加快加载速度并提高连接稳定性。

Each step in the internet’s capabilities led to a new investment in CDN infrastructure. DDoS and cyberwarfare put CDNs on the physical frontline, originally with physical firewalls. Video streaming prompted increased bandwidth and capacity of CDN networks to meet booming consumer demand.互联网功能的每一次提升都促使人们对 CDN 基础设施进行新的投资。DDoS 攻击和网络战使 CDN 成为抵御网络攻击的物理前线，最初需要借助物理防火墙。视频流媒体的兴起则促使 CDN 网络带宽和容量的提升，以满足消费者日益增长的需求。

Each of these waves eventually enters a commoditization cycle, and share prices still reflect that recent carnage. Pricing wars turned the space into a graveyard: StackPath and Lumen exited the business in 2023-24 while Edgio filed for bankruptcy protection in September 2024.每一波浪潮最终都会进入商品化周期，而股价至今仍反映着近期的惨淡局面。价格战将这个领域变成了坟场：StackPath 和 Lumen 在 2023-2024 年退出了该行业，而 Edgio 则在 2024 年 9 月申请了破产保护。

The surviving firms with location-advantaged infrastructure are becoming increasingly valuable providers of low-latency AI compute. The coming surge of agentic traffic should further solidify this market, as “CDN” transforms from Content Delivery Networks to Compute Delivery Networks.那些拥有地理位置优势基础设施的幸存企业正日益成为低延迟人工智能计算领域的重要提供商。即将到来的智能体流量激增将进一步巩固这一市场，因为“CDN”正从内容分发网络转型为计算分发网络。

Akamai (AKAM US) stands to benefit from a double pivot – repurposing its content delivery footprint for AI compute, while also rapidly expanding its cybersecurity business.Akamai (AKAM US) 有望从双重转型中受益——将其内容分发业务重新用于 AI 计算，同时迅速扩展其网络安全业务。

The cyber side is already seeing tailwinds from AI security as both agents and applications are hitting APIs hard. The demand for API security has been explicitly highlighted by AKAM management as a large tailwind, with revenue for this segment growing 100% YoY.网络安全领域已从人工智能安全中受益，因为智能体和应用程序都在大量使用 API。AKAM 管理层明确指出，对 API 安全的需求是一大利好因素， 该业务板块的收入同比增长了 100%。

The company is now leveraging their existing CDN footprint to build out AI services. In October 2025, Akamai launched the Akamai Inference Cloud, bolstered by $250 million in capex tied to their cloud buildout. Investors panicked following the company’s earnings release, although concerns were largely tied to opex discipline, rather than any forecasted drop in demand.该公司目前正利用其现有的 CDN 网络构建人工智能服务。2025 年 10 月，Akamai 推出了 Akamai 推理云 ，并为此投入了 2.5 亿美元的资本支出用于云建设。尽管投资者对该公司财报感到恐慌，但他们的担忧主要集中在运营支出控制方面，而非对需求下降的预期。

A similar turnaround is rapidly unfolding at Fastly (FSLY US). The company was left for dead following its run in 2021, which we feel gives it the highest convexity among CDN plays. Agentic traffic helped push the company into positive free cash flow territory, with GAAP losses being driven heavily by SBC dilution.Fastly (FSLY US) 也正在迅速实现类似的逆转。该公司在 2021 年的强劲增长之后一度被认为已经走向衰亡，我们认为这使其在 CDN 服务商中具有最高的凸性。代理流量帮助该公司扭亏为盈，而 GAAP 亏损则主要源于 SBC 的股权稀释。

In February 2026, the company rolled out an AI Accelerator, where they provide semantic caching that decreases both latency and inference costs. This offering complements their Compute platform, using WebAssembly (WASM) as the foundation. This choice signals that they are attempting to live up to their brand name: optimizing for speed surrounding high-density networks instead of trying to maximize overall network coverage.2026 年 2 月，该公司推出了一款人工智能加速器 ，提供语义缓存功能，可降低延迟和推理成本。该产品是对其计算平台的补充，并以 WebAssembly (WASM)为基础。这一选择表明，他们正努力践行其品牌理念：专注于在高密度网络环境下优化速度， 而非追求最大化网络覆盖范围。

API GatewaysAPI 网关

This CDN build-out creates a subtle problem in that getting traffic to the edge faster is only useful if what’s waiting at the other end can handle it. Every API call an agent makes passes through a load balancer before it ever reaches a backend application.这种 CDN 构建方式带来了一个微妙的问题：只有当后端服务器能够处理这些流量时，更快地将流量传输到边缘才有意义。代理发出的每个 API 调用在到达后端应用程序之前都会先经过负载均衡器。

F5 Inc (FFIV US) owns NGINX, the most widely deployed reverse proxy on the internet. Their NGINX One platform unifies all NGINX deployments into a single control plane, and management has been explicit in their repositioning of it as the API orchestration layer for agentic traffic. On their Q1 FY2026 earnings call, the CFO noted that AI-generated API call volumes are running “meaningfully ahead” of prior projections, with non-human traffic now the faster-growing segment of throughput metrics.F5 公司（FFIV US）拥有 NGINX，它是互联网上部署最广泛的反向代理。其 NGINX One 平台将所有 NGINX 部署统一到一个控制平面中，管理层已明确表示将其重新定位为代理流量的 API 编排层。在 2026 财年第一季度财报电话会议上，首席财务官指出，人工智能生成的 API 调用量“显著超出”此前的预期，非人类流量现在已成为吞吐量指标中增长最快的部分。

Data Center REITs & Cloud Providers数据中心 REITs 和云服务提供商

While perhaps not the sexiest category, we also think that data center REITs should continue to benefit from the tailwinds of both the shape and latency needs of agentic AI.虽然数据中心 REITs 可能不是最吸引人的类别，但我们也认为，随着智能 AI 的形状和延迟需求的变化，数据中心 REITs 将继续受益。

Equinix (EQIX US) owns the buildings that hold the data centers, with about 70% of revenue attributed to this business line. The bull case is about as first-order as you can get: more data centers leads to more colocation demand, which drives more sales and higher rents.Equinix（EQIX US）拥有数据中心所在的建筑，该业务贡献了公司约 70%的收入。其利好前景显而易见：更多的数据中心会带来更多的托管需求，进而推动销售额增长和租金上涨。

The convexity comes from its Interconnect revenue. This is a recurring service that Equinix offers to connect servers – effectively building extra lanes on the east-west highway. They offer physical connects as well as a virtual interconnect layer called Equinix Fabric, which allows “interconnect on demand.” The company recently reported 500,000 interconnects and 60% of their large new contracts are from AI workloads. Digital Realty Trust (DLR US) is the other long-standing DC REIT with similar positioning to EQIX. DLR is rapidly closing the gap in high-margin connectivity via their ServiceFabric offering, a software-defined orchestration layer built to rival Equinix’s Fabric product.Equinix 的凸性优势源于其互连业务收入。这是一项 Equinix 提供的服务器连接服务，相当于在东西向高速公路上增设了车道。他们提供物理连接以及名为 Equinix Fabric 的虚拟互连层，后者支持“按需互连”。该公司近期报告称，其互连数量已达 50 万个，且 60% 的大型新合同来自人工智能工作负载。DigitalRealty Trust (DLR US) 是另一家与 Equinix 定位相似的、历史悠久的华盛顿特区房地产投资信托基金 (REIT)。DLR 正通过其 ServiceFabric 产品迅速缩小与 Equinix 高利润互连业务的差距。ServiceFabric 是一个软件定义的编排层，旨在与 Equinix 的 Fabric 产品竞争。

Overseas, we highlight VNET Group (VNET US) which operates as a carrier- and cloud-neutral internet data center services provider in China. We see China leading in real-world agentic adoption but lagging the US on compute buildout. VNET has shown significant business reacceleration in the past two years that has few signs of slowing.在海外，我们重点关注 VNET 集团（VNET US）， 该公司在中国运营，是一家运营商和云中立的互联网数据中心服务提供商。我们认为中国在实际应用智能体方面处于领先地位，但在计算资源建设方面落后于美国。VNET 在过去两年中业务增长显著，且目前尚无放缓迹象。

On the splashier side, the market has been rapidly re-rating DigitalOcean (DOCN US). Where hyperscale cloudswin business on up-time, price, and peace-of-mind, DigitalOcean acts as more of a high-touch partner rather than a DIY vendor.另一方面，市场正在迅速重新评估 DigitalOcean (DOCN US)。 超大规模云服务凭借正常运行时间、价格和安心保障赢得业务，而 DigitalOcean 则更像是一位提供高度个性化服务的合作伙伴，而非自助式供应商。

While this may lead to slightly-higher unit costs of compute, the value-add services that DigitalOcean provides for small- and medium-sized enterprises and startups tend to drive cost savings in other areas of the budget, say in compliance, security, and even financial planning. Larger clouds like AWS and Azure claim to offer similar support, but good luck getting world-class service if your company doesn’t have an eight-figure-plus bill at the end of each month.虽然这可能会导致计算成本略微上升，但 DigitalOcean 为中小企业和初创公司提供的增值服务往往能帮助企业在预算的其他方面节省成本，例如合规性、安全性，甚至财务规划。像 AWS 和 Azure 这样的大型云平台声称提供类似的支持，但如果你的公司每月账单没有达到八位数以上，想要获得世界一流的服务就难上加难了。

Managed-service providers like DigitalOcean exist in many industries to lower barriers to entry and decrease friction for emerging businesses. With the velocity of ideas throughout the economy reaching a fever pitch, DigitalOcean provides a platform that greases the wheels between idea and implementation.像 DigitalOcean 这样的托管服务提供商遍布各行各业，旨在降低新兴企业的准入门槛，减少其发展阻力。随着经济领域内创意涌现速度飞快，DigitalOcean 提供了一个平台，能够有效促进创意与实践之间的衔接。

This shift towards digital native enterprises (DNEs), a fancy term for AI/agentic startups, has resulted in a dramatic re-acceleration in the topline, and surprisingly compelling cash flow vs. other neocloud platforms.这种向数字原生企业（DNE，人工智能/智能体初创公司的一个时髦术语）的转变，已经导致营收大幅回升，并且与其他新云平台相比，现金流也出人意料地强劲。

Another interesting, if more risky, play is Rackspace (RXT US). Until recently, it appeared to be following the Wolfspeed (WOLF US) playbook: a debt-laden company brought public by Apollo with the equity heading towards a full writedown.另一个有趣但风险更高的投资标的是 Rackspace (RXT US)。直到最近，它似乎都在效仿 Wolfspeed (WOLF US) 的模式：一家负债累累的公司，由 Apollo 将其上市，而其股权正面临全面减记。

The company managed a stick save in February, inking a partnership with Palantir, where Rackspace would provide hosting and systems integration services (the latter of which being a new revenue category). Of course, there’s still a risk that their core hosting business continues to shrink, even with their new agentic offerings.今年二月，该公司成功挽救了颓势，与 Palantir 达成合作，Rackspace 将为其提供主机托管和系统集成服务（后者是一个新的收入来源）。当然，即便推出了新的代理服务，其核心主机托管业务仍面临持续萎缩的风险。

The stock offers high convexity to agentic demand, though the leverage is still there with $2.75 billion in long term debt.该股票对代理需求具有很高的凸性，但其长期债务高达 27.5 亿美元，杠杆率仍然很高。

The Agentic Conglomerate: Cloudflare代理集团：Cloudflare

Cloudflare (NET US) is evolving into the most comprehensive play on agentic demand. While they have the structure of a traditional CDN, they use that footprint to sell services on top of their infrastructure.Cloudflare（美国）正在发展成为代理需求领域最全面的解决方案提供商。虽然其架构与传统 CDN 类似，但它利用这一基础设施之上的服务来销售内容分发服务。

Those services fit agents like a glove. They released Workers in 2017 that allow users to run operations on Cloudflare’s distributed network rather than on a single server in Northern Virginia. Five years later they added Durable Objects, which provided a “state” to their Workers.这些服务与代理商完美契合。他们在 2017 年发布了 Workers，使用户能够在 Cloudflare 的分布式网络上运行操作，而不是在弗吉尼亚州北部的一台服务器上运行。五年后，他们又添加了 Durable Objects，为 Workers 提供了一种“状态”管理。

When agentic demand hit, the scaffolding was already there.当市场需求出现时，一切准备工作已经就绪。

Cloudflare is also a first mover in agentic ecommerce. They created NET dollar, a stablecoin for agentic transactions. They co-founded the x402 Foundation with Coinbase, and are working with major payment companies to develop the authentication layer for agentic ecommerce.Cloudflare 也是代理电子商务领域的先行者。他们创建了 NET Dollar，一种用于代理交易的稳定币。他们与 Coinbase 共同创立了 x402 基金会 ，并正与各大支付公司合作，开发代理电子商务的身份验证层。

Finally, the company also is a critical governance player, providing a unified security layer that governs how autonomous AI agents interact with corporate data and external APIs. By integrating tools like Firewall for AI and AI Audit, Cloudflare enables organizations to enforce least-privilege access and real-time guardrails, ensuring that agentic workflows remain secure and compliant without stifling their autonomy.最后，该公司也是重要的治理参与者，提供统一的安全层，用于管理自主 AI 代理与企业数据和外部 API 的交互方式。通过集成 Firewall for AI 和 AI Audit 等工具，Cloudflare 帮助企业实施最小权限访问和实时防护措施，确保代理工作流程安全合规，同时又不限制其自主性。

But you shouldn’t imagine agents living in clean APIs and structured tool environments. Real world enterprises run on terrible web UIs, fragile workflows, vendor portals with session timeouts & systems that only work because a human is squinting at a screen clicking through four confirmation dialogs. So, agents will need browser infrastructure the same way knowledge workers need laptops.但你不应该想象智能体生活在干净的 API 和结构化的工具环境中。现实世界的企业运行在糟糕的 Web 用户界面、脆弱的工作流程、带有会话超时的供应商门户以及只有人工眯着眼睛盯着屏幕点击四个确认对话框才能正常运行的系统上。因此，智能体需要浏览器基础设施，就像知识工作者需要笔记本电脑一样。

Cloudflare is already pushing Browser Rendering, signed agent traffic, remote MCP plumbing, and something they call “Markdown for Agents,” which is basically the web learning to speak machine-first. Zscaler (ZS US) just acquired SquareX to secure browsers for the AI era. Of these, Cloudflare is doing the most interesting work.Cloudflare 已经在大力推广浏览器渲染、签名代理流量、远程 MCP 管道以及他们所谓的“代理 Markdown”，这项技术本质上是让 Web 学习如何优先与机器交互。Zscaler (ZS US) 刚刚收购了 SquareX，旨在为人工智能时代保障浏览器安全。在这些公司中，Cloudflare 的工作最为引人注目。

The “bear case” is that NET’s strong positioning in the agentic landscape is well known and already baked into today’s valuation, unlike some of the “turnaround” stories mentioned prior. Nevertheless, we think that owning high quality, highly exposed names is worth the price of admission.看跌的理由是，NET 在代理领域的强势地位众所周知，并且已经反映在目前的估值中，这与之前提到的一些“扭亏为盈”的故事截然不同。尽管如此，我们认为持有高质量、高曝光度的公司是值得的。

2) Ecosystem2）生态系统

If infrastructure provides agents with the digital foundation to perform human-like workflows, then the ecosystem brings together the tools that agents will use to work with humans. After all, the objective of this innovation is to harness agentic capabilities to improve real-world outcomes and solve for human inefficiency.如果基础设施为智能体提供了执行类似人类工作流程所需的数字基础，那么生态系统则汇集了智能体与人类协作所需的各种工具。毕竟，这项创新的目标是利用智能体的能力来改善现实世界的结果，并解决人类效率低下的问题。

SaaS businesses are not blind to the existential risks facing their business models. To contend against this looming threat, every software company under the sun is coming to market with their own agentic offering. Of course, some (if not most) of these products are too little, too late.SaaS 企业并非对自身商业模式面临的生存风险视而不见。为了应对这一迫在眉睫的威胁，几乎所有软件公司都纷纷推出各自的代理产品。当然，其中一些（即便不是大多数）产品推出得太晚，也远远不够。

On the other hand, the “moat” in software is often unrelated to the sophistication of the technology itself. Software companies with valuable data silos, physical infrastructure, and regulatory advantages are likely to succeed in an agentic world. The companies that function as a toll on agent-to-human interaction are harder to “vibecode” away.另一方面，软件领域的“护城河”往往与技术本身的复杂程度无关。 拥有宝贵数据孤岛、物理基础设施和监管优势的软件公司更有可能在智能体主导的世界中取得成功。 那些阻碍智能体与人之间互动的公司则更难被“氛围编码”式的手段淘汰。

After all, every human-agent interaction must begin somewhere: A human prompts an agent. An agent then accesses an API, pays for something it will use, dispatches a phone call, email blast, or iMessage to the end user or any number or permutation of tasks.毕竟，每一次人机交互都必须从某个地方开始：用户向智能体发出指令。然后，智能体访问 API，支付所需资源的费用，向最终用户发送电话、电子邮件群发或 iMessage 消息，或者执行任何其他数量或组合的任务。

Payments付款

We spoke about the reasoning behind stablecoins as an agentic payment rail in our “2028 Global Intelligence Crisis” piece. We find agentic payment rails and agentic commerce very interesting, because they mean that payment rails can actually compete. When consumers get a card from the bank, they don’t care whether it’s Mastercard or Visa, they just use it. This can change significantly with AI agents making switching between rails more dynamic.我们在题为《2028 年全球情报危机》的文章中探讨了稳定币作为智能支付渠道背后的逻辑。我们认为智能支付渠道和智能商务非常有趣，因为它们意味着不同的支付渠道之间可以真正展开竞争。当消费者从银行拿到一张卡时，他们并不关心是万事达卡还是维萨卡，他们只会使用它。但随着人工智能代理的出现，支付渠道之间的切换将变得更加动态，这种情况可能会发生显著变化。

After all, we still live in a capitalist society – agents are much less useful if they’re unable to transact on our behalf. Pulling out your credit card every time you make a purchase online feels like a prehistoric practice – there must be some sort of improvement.毕竟，我们仍然生活在资本主义社会——如果代理商无法代表我们进行交易，他们的作用就大打折扣。每次网上购物都要掏出信用卡，感觉就像是史前时代的做法—— 必须有所改进才行 。

These agentic tailwinds come at a moment when stablecoin adoption has already seen dramatic growth as regulatory clarity has allowed for institutional buy-in.在这些推动稳定币普及的顺风之下，监管的明朗化使得机构投资者能够参与其中，稳定币的采用率已经出现了显著增长。

Early agentic adopters have discovered that it’s much easier to provision an agent with a crypto wallet than with a credit card. That has opened up an entire sector-wide effort to create agentic payment workflows.早期采用代理支付技术的用户发现，为代理配置加密钱包比配置信用卡要容易得多。这促使整个行业致力于创建代理支付工作流程。

According to Visa’s onchain analytics dashboard, we see USDC transaction volume excluding crypto exchanges surging post the passing of the GENIUS Act in July 2025.根据 Visa 的链上分析仪表板显示，在 2025 年 7 月 GENIUS 法案通过后，不包括加密货币交易所的 USDC 交易量将激增。

We rarely comment on crypto, with the exception of our well-known bullishness on Hyperliquid as they integrate 24/7 markets for commodities, FX, and stocks. However, it’s important to understand what’s being built to enable agentic commerce if it’s going to impact stocks.我们很少评论加密货币，但众所周知，我们非常看好 Hyperliquid，因为他们整合了全天候的商品、外汇和股票市场。然而，如果代理交易要对股票市场产生影响，那么了解其背后的运作机制就至关重要。

x402

If you’ve ever seen the error “404 not found”, then you have experience with an HTTP status code. There’s a lot more from this category, including “301 – redirect”, “500 – internal service error”, and “403 – forbidden”. While all different, they’re all incredibly frustrating to see on your computer screen.如果你曾经见过 “404 未找到” 错误，那么你对 HTTP 状态码就有所了解。这个类别下还有很多其他状态码，例如“301 重定向”、“500 内部服务错误” 和“403 禁止访问”。虽然它们各不相同，但看到它们出现在电脑屏幕上都令人非常沮丧。

Some of the response codes have been reserved, but are not commonly found. One instance of note is “402 – payment required”, a response that was once reserved for a future use case.有些响应代码已被保留，但并不常用。例如，“402 – 需要付款” 就是一个值得注意的例子，该响应代码曾被保留用于未来的特定用途。

That use case is now here.现在，这个应用场景已经出现了。

The similarities between HTTP and agentic payment protocols are significant. We feel that understanding agentic payment protocols is important to one day understand the winners of it, but it’s equally important to understand that value won’t accrue to the protocol layer that enables agents (rather, they will accrue to APIs in demand, agentic harnesses and the agents themselves). So, with that being said, let’s look at one of the protocols – x402.HTTP 和代理支付协议之间的相似之处非常显著。我们认为，理解代理支付协议对于日后了解其最终赢家至关重要，但同样重要的是要明白，价值不会归于支持代理的协议层（而是归于需求旺盛的 API、代理框架以及代理本身）。因此，接下来让我们来看其中一个协议——x402。

The “x402 open standard” is a crypto-native payments standard that uses the existing hypertext protocol for agentic transactions. x402 was jointly developed by Coinbase (COIN US) and is intended to serve as the payments protocol for AI agents, Cloudflare (NET US) has also joined the foundation responsible for its development and implementation. It joins“x402 开放标准” 是一种加密原生支付标准，它利用现有的超文本协议进行智能体交易。x402 由 Coinbase (COIN US) 和 Cloudflare (NET US) 联合开发，旨在作为人工智能代理的支付协议。Cloudflare 也加入了负责 x402 开发和实施的基金会。

In July 2025, Cloudflare developed “pay per crawl,” allowing content owners to charge a fee for an AI to access its content. A few months later, Cloudflare announced NET Dollar, a USD-backed stablecoin built for agentic commerce. It’s clear that NET is positioning itself across agentic infrastructure to be a clear cut winner.2025 年 7 月，Cloudflare 推出了“按抓取付费”模式，允许内容所有者向人工智能收取访问其内容的费用。几个月后，Cloudflare 发布了 NET Dollar ， 这是一种与美元挂钩的稳定币，专为智能体商务而构建。显然，NET 正致力于在智能体基础设施领域占据一席之地，力争成为行业翘楚。

Circle Internet Group (CRCL US) has released Circle Nanopayments, which “enables gas-free USDC transfers as small as $0.000001”, positioning the x402 protocol to serve as the payment rail for agentic transactions, abetted by the proliferation of stablecoin acceptance and infrastructure.Circle Internet Group (CRCL US) 发布了 Circle Nanopayments ，该技术 “支持低至 0.000001 美元的免 gas 费用 USDC 转账”，使 x402 协议能够作为代理交易的支付渠道，这得益于稳定币的普及和基础设施的完善。

“[…]agent-to-agent transactions need a reliable, low cost, trusted medium of exchange. And so virtually, all of the AI payments infrastructure that we’re seeing, the agent-to-agent type activity is happening with blockchains, it is happening with USDC. ““[…]代理之间的交易需要可靠、低成本、值得信赖的交易媒介。因此，我们看到的几乎所有人工智能支付基础设施，以及代理之间的交易活动，都是通过区块链或 USDC 实现的。”

– Circle Internet Group Q4 2025 Earnings Call– Circle Internet Group 2025 年第四季度财报电话会议

While the stock has been relegated to the doldrums of Value Investor Club threads, PayPal (PYPL US) has also issued a USD-backed stablecoin – PYUSD, whose circulation grew significantly in 2H25, reaching over $4 billion. While still far, far behind USDC and USDT (Tether) in terms of market adoption, we view PYPL as a longer-shot stablecoin play given its single-digit earnings multiple.尽管 PayPal (PYPL US) 的股票在价值投资者俱乐部论坛上一直处于低迷状态，但该公司也发行了一种以美元为支撑的稳定币 ——PYUSD。PYUSD 在 2025 年下半年的流通量显著增长，超过 40 亿美元。虽然 PYUSD 在市场接受度方面仍远远落后于 USDC 和 USDT（Tether），但考虑到其个位数的市盈率，我们认为 PYPL 是一项值得长期投资的稳定币。

To get more in depth on why those micro transactions are necessary, let’s look at another protocol being developed for agentic payments:为了更深入地了解为什么这些微交易是必要的，让我们来看看另一个正在开发的代理支付协议：

Tempo & MPP时间和 MPP

Backed by Stripe, Paradigm and Visa, Tempo is attempting to solve how agents manage their balance sheet continuously in order to complete a task using an L1 blockchain explicitly designed for agentic stablecoin transactions. At the highest level of abstraction, you can think of MPP as a way to make payments a first class citizen in the HTTP request response cycle. MPP is designed to be an IETF approved standard that works with any payment method. Leveraging alliances with incumbents like Visa and Lightspark, MPP facilitates settlement across both legacy credit card infrastructure and crypto as a “rail-agnostic” primitive for the agentic economy.在 Stripe、Paradigm 和 Visa 的支持下，Tempo 致力于解决代理商如何持续管理其资产负债表以完成任务的问题，其使用的 L1 区块链专为代理商稳定币交易而设计。从最高抽象层面来看，您可以将 MPP 理解为一种使支付成为 HTTP 请求响应周期中“一等公民”的方式。MPP 旨在成为 IETF 认可的标准，并可与任何支付方式兼容。凭借与 Visa 和 Lightspark 等现有支付巨头的合作，MPP 作为代理商经济的“与支付系统无关”的基础机制，促进了传统信用卡基础设施和加密货币之间的结算。

In a traditional workflow, payments are discrete and user-initiated. In an agentic workflow, they are continuous, conditional, and embedded inside decision trees. Tempo abstracts this into a programmable “payment clock”, allowing agents to stream value, escrow funds, or dynamically re-route spend based on real-time outcomes. The Machine Payments Protocol extends this further by standardizing how agents negotiate price, authorize spend, and settle transactions across counterparties without human intervention. What does this mean?在传统工作流程中，支付是离散的，由用户发起。而在代理工作流程中，支付是连续的、有条件的，并且嵌入在决策树中。Tempo 将其抽象为一个可编程的“支付时钟”，允许代理根据实时结果流式传输价值、托管资金或动态重新分配支出。机器支付协议进一步扩展了这一功能，它标准化了代理如何在无需人工干预的情况下与交易对手协商价格、授权支出和结算交易。 这意味着什么？

In layman’s terms, you tell an agent you want to spend no more than $5 on a task – it goes out and draws from that $5. Essentially, Tempo allows a cryptographically signed message to authorize an amount held in non-custodial escrow, which an agent can incrementally debit. The reasoning behind this is that agents can potentially require thousands of micropayments per task. The human pays once, the agent pays thousands of times (without latency or waiting for block confirmations). You can see how this would be impractical with a credit card, where transaction costs would almost immediately exceed the actual value of transactions.简单来说，你告诉代理你希望完成一项任务的花费不超过 5 美元——它就会从这 5 美元中扣款。本质上，Tempo 允许通过加密签名消息授权在非托管账户中预留一笔资金，代理可以分批从中扣款。这样做的理由是，代理可能需要为每个任务支付数千笔小额款项。用户只需支付一次，代理就可以支付数千次（无需延迟或等待区块确认）。不难想象，如果使用信用卡，这种方式会多么不切实际，因为交易成本几乎会立即超过交易的实际金额。

This is a topic deserving of its own primer, probably written by someone more crypto-competent than we are, but there will be a plethora of problems and solutions proposed inherent to how machines will pay each other.这是一个值得专门撰写入门指南的话题，或许应该由比我们更精通加密货币的人来撰写，但其中必然包含着大量关于机器之间如何支付的固有问题和解决方案。

For example, what happens when a wallet has stablecoins but runs out of the native blockchain’s token and doesn’t have enough gas for the transaction? For the uninitiated, in order to send USDC on a blockchain such as Ethereum, you also need a nominal amount of ETH to pay for the transaction. That can interrupt a workflow and result in needing human involvement. Those Circle nanopayments mentioned earlier do not work on Ethereum or Solana chains, because the chains themselves charge for gas.例如，当钱包里有稳定币，但本地区块链的代币用完了，且没有足够的 gas 费进行交易时会发生什么？对于不熟悉的人来说，要在以太坊等区块链上发送 USDC，还需要支付一定数量的 ETH 作为交易费用。这可能会中断工作流程，并导致需要人工干预。前面提到的 Circle 纳米支付在以太坊或 Solana 链上无法使用，因为这些链本身会收取 gas 费。

One example of a solution is Stable, a blockchain built for stablecoin payments. Stable’s single-token architecture directly addresses this by using USDT, the world’s largest stablecoin, for both gas and settlement. In this model, an agent’s cost accounting becomes fully deterministic in dollar terms with no gas conversion spread, no exposure to native token volatility, and no need to maintain a separate balance solely to execute transactions.Stable 就是一个解决方案的例子，它是一个专为稳定币支付而构建的区块链。Stable 的单代币架构直接解决了这个问题，它使用全球最大的稳定币 USDT 作为 gas 费用和结算货币。在这种模式下，交易者的成本核算完全以美元计价，无需支付 gas 转换费，无需承担原生代币价格波动的风险，也无需为了执行交易而维护单独的余额。

Notably, this is just one pillar of “agentic commerce”. Again, the tech is still in its skeuomorphicphase – we expect agents to transact like a human might. Looking ahead, don’t be shocked when there are agent-to-agent (A2A) marketplaces where prices are dynamically quoted with little human input.值得注意的是，这只是“智能体商务”的一个支柱。再次强调，这项技术仍处于拟物化阶段。现阶段——我们期望代理商像真人一样进行交易。展望未来，当出现价格几乎无需人工干预、动态生成的代理商对代理商（A2A）市场时，请不要感到惊讶。

It’s worth a caveat that this is all extremely early. It seems trivial to predict that agents will have to pay for things, and almost assuredly correct, but there’s a long way to go both in terms of tech and adoption.需要指出的是，这一切都还处于非常早期的阶段。预测经纪人将来需要付费似乎理所当然，而且几乎可以肯定是正确的，但无论从技术层面还是普及程度来看，还有很长的路要走。

It will be top of mind for us to keep track of how some of the bigger players in payments (like V, MA, GPN, FIS, FISV, SHOP etc.) integrate and interact with agentic payments. All of these companies have at least mentioned agentic commerce. Fiserv, in January, announced their platform meant to let banks identify and authorize AI-initiated transactions while adding fraud protection and preserving top-of-wallet positioning – attempting to position themselves as a picks-and-shovels play for banks that do not wish to be disintermediated when software starts shopping.我们将密切关注一些大型支付公司（例如 V、MA、GPN、FIS、FISV、SHOP 等）如何整合代理支付并与之互动。所有这些公司都至少提及过代理商务。Fiserv 在 1 月份发布了其平台，旨在帮助银行识别和授权人工智能发起的交易，同时增加欺诈保护并保持其在客户钱包中的首选地位——试图将自己定位为那些不希望在软件开始“抢占市场”时被银行取代的“得力助手”。

Still, while important to understand, we feel it’s far too early to begin investing in payments solely on agentic upside before these companies truly demonstrate their prowess and ability to adapt.不过，虽然了解这一点很重要，但我们认为，在这些公司真正展现出实力和适应能力之前，仅仅因为代理机构的潜在收益就开始投资支付领域还为时过早。

Last Mile Telephony最后一公里电话

According to the Communication Workers of America (CWA), 3.6 million people are employed by call centers, an estimated 2.5% share of the US workforce. At $15/hour, we estimate that the US spends more than $100 billion annually on call center labor.根据美国通信工人协会（CWA）的数据，美国有 360 万人受雇于呼叫中心，约占美国劳动力的 2.5%。按每小时 15 美元的工资计算，我们估计美国每年在呼叫中心劳动力上的支出超过 1000 亿美元。

The magnitude of this expense has lured many software businesses into the call center as a service, or CCaaS marketplace. However, these businesses are not immune to the headwinds facing the software ecosystem. An agentic competitor or a savvy voice AI startup can seemingly disrupt an incredible product in the blink of an eye.如此巨大的成本吸引了众多软件企业涌入呼叫中心即服务 （CCaaS）市场。然而，这些企业也无法免受软件生态系统所面临的逆境的影响。一个咄咄逼人的竞争对手或一家精明的语音人工智能初创公司，似乎就能在一瞬间颠覆一款优秀的产品。

An agentic voice agent, however, cannot effectively replicate telephone infrastructure on its own. You need public switched telephone network (PSTN) interconnects, carrier licenses, local number portability, and actual physical infrastructure in order to get a phone number.然而，语音代理本身无法有效地复制电话基础设施。要获得电话号码，您需要公共交换电话网络 (PSTN) 互连、运营商许可证、本地号码可携性以及实际的物理基础设施。

That is the business that Bandwidth (BAND US) operates. The world’s largest technology companies – Microsoft, Google, Amazon’s AWS – all pay Bandwidth for access to their telephone infrastructure, signaling that it’s easier to pay this toll to Bandwidth than building out your own telephone infrastructure.这就是 Bandwidth（BAND US） 的业务。全球最大的科技公司——微软、谷歌、亚马逊的 AWS——都向 Bandwidth 付费以使用其电话基础设施，这表明向 Bandwidth 支付这笔费用比自行建设电话基础设施要容易得多。

In September 2025, Bandwidth announced full integration into OpenAI’s Realtime API, allowing enterprises to “bring their own AI” and quickly spin up conversational voice agents using Bandwidth’s edge infrastructure.2025 年 9 月，Bandwidth 宣布与 OpenAI 的 Realtime API 完全集成，使企业能够 “自带 AI”，并使用 Bandwidth 的边缘基础设施快速启动对话式语音代理。

“Enterprises can now deploy AI voice agents powered by the technology behind ChatGPT, with the reliability and global scale of our Communications Cloud.“企业现在可以部署由 ChatGPT 背后的技术驱动的 AI 语音代理，并利用我们通信云的可靠性和全球规模。

By offering many different options to integrate conversational AI, Bandwidth is becoming the AI orchestration leader for the global enterprise.”Bandwidth 提供多种不同的对话式 AI 集成方案， 正成为全球企业 AI 编排领域的领导者。

– John Bell, Bandwidth Chief Product Officer– John Bell，带宽首席产品官

Agentic communication, of course, will not be limited to phone calls. Knowing the communication patterns of Gen Z consumers – text will be the predominant interface of agent-to-human interaction. This is already evident from the popularity of LLM chatbots. However, without proper context, agents fail to address the acute needs of the customer while risking churn and missed opportunities in re-selling, upselling, or retaining customers.当然，智能体沟通不会局限于电话。了解 Z 世代消费者的沟通模式后，我们发现文字将成为智能体与人互动的主要界面 。LLM 聊天机器人的流行已经印证了这一点。然而，如果缺乏恰当的上下文信息，智能体就无法满足客户的迫切需求，反而可能导致客户流失，并错失再次销售、追加销售或客户维系的机会。

Twilio (TWLO US) rounds out the communication layer for agents with Conversational Intelligence. While Bandwidth largely tackles the customer service “voice” angle of AI communications, Twilio’s edge hails from fifteen years of customer data embedded within their Segment product. This context-rich corpus of data empowers Twilio’s customers to harness agentic AI – across voice AI and text – to understand their customers and achieve optimal sales outcomes.Twilio (TWLO US) 通过对话智能完善了客服人员的通信层。虽然 Bandwidth 主要着眼于人工智能通信中客户服务的“语音”方面，但 Twilio 的优势在于其 Segment 产品中嵌入了十五年的客户数据。这些包含丰富上下文信息的数据集使 Twilio 的客户能够利用智能代理人工智能（涵盖语音和文本）来了解客户并实现最佳销售业绩。

Ironically, Twilio’s Segment acquisition was lambasted by the street as a huge overpay in the heat of 2020-21 valuations. Activists lobbied for Twilio to spin out the Segment division in 2023-24 as it dragged on profit margins and detracted from the company’s core business model. Founder and CEO Jeff Lawson ultimately stepped down amidst this activist quarrel, although the Segment division was never sold. This acquisition, while expensive at the time, may prove to be prescient.具有讽刺意味的是，在 2020-2021 年估值高峰期，Twilio 收购 Segment 的交易被华尔街抨击为溢价过高。激进投资者游说 Twilio 在 2023-2024 年剥离 Segment 部门，因为该部门拖累了利润率，并偏离了公司的核心业务模式。创始人兼首席执行官杰夫·劳森最终在这场激进投资者的纷争中辞职，尽管 Segment 部门从未被出售。尽管当时的收购价格昂贵，但事实证明这笔交易极具前瞻性。

In May 2025, Twilio reached a pivotal milestone with the general availability of ConversationRelay, a product designed for developers to build human-like AI voice agents. ConversationRelay can be viewed as an “out-of-the-box” offering for agentic voice AI – bundling the fragmented functions of speech-to-text, analytics, and audio – all into one product.2025 年 5 月，Twilio 正式发布了 ConversationRelay，这标志着其发展历程中的一个重要里程碑。ConversationRelay 是一款专为开发者打造类人 AI 语音代理而设计的产品。它可以被视为一款“开箱即用”的智能语音 AI 解决方案，将语音转文本、分析和音频等分散功能整合到一个产品中。

Both Bandwidth and Twilio’s products are HIPAA compliant, meaning that medical records can be transmitted using programmable SMS and Voice. This too, is an enormous regulatory advantage as healthcare customers are extremely sensitive about sharing patient information. LLMs remain somewhat unreliable from a cybersecurity standpoint, but we’ll touch upon this later.Bandwidth 和 Twilio 的产品均符合 HIPAA 标准，这意味着可以通过可编程短信和语音传输医疗记录。 这对于医疗保健客户而言也是一项巨大的监管优势 ，因为他们对共享患者信息极其敏感。从网络安全角度来看，LLM 的可靠性仍然存在一些问题，我们稍后会详细讨论这一点。

Of course, the risk of frontier models emulating telephony solutions still remains. However, a major advantage of both Bandwidth and Twilio is at the orchestration layer: These companies have long-standing relationships with carriers that manage interconnects and regulatory compliance. There are deep integrations with phone systems and existing enterprise partners wherein these features can be quickly accretive to the company.当然，新兴模式模仿电话解决方案的风险依然存在。然而，Bandwidth 和 Twilio 的一大优势在于编排层： 这两家公司与运营商建立了长期合作关系，运营商负责管理互联互通和合规事宜。它们与电话系统和现有企业合作伙伴实现了深度集成，这些功能可以迅速为公司带来收益。

“[…]we are moving beyond being a provider of communications channels and data toward becoming a foundational infrastructure layer in the age of AI. Revenue from our voice channel continues to accelerate, aided in part by voice AI, which we believe is just the beginning, as these use cases will evolve to be more conversational and cross-channel, an area where Twilio is uniquely differentiated.”“我们正在从通信渠道和数据提供商转型为人工智能时代的基础架构层 。语音渠道的收入持续增长，这在一定程度上得益于语音人工智能，我们相信这仅仅是个开始，因为这些用例将发展得更具对话性和跨渠道性，而这正是 Twilio 的独特优势所在。”

– Khozema Shipchandler, Twilio CEO– Twilio 首席执行官 Khozema Shipchandler

Like Bandwidth, Twilio is fully integrated into OpenAI’s Realtime API. Twilio itself already handles 85% of its own inbound sales calls through AI agents, proving that this model can work.与 Bandwidth 一样，Twilio 也完全集成到 OpenAI 的实时 API 中 。Twilio 自身已经通过 AI 代理处理了 85% 的来电销售咨询，这证明了这种模式的可行性。

Last month, Twilio launched the Agent-to-Human (A2H) protocol, designed to complement MCP and Agent-to-Agent (A2A) architectures by standardizing how AI agents interface with humans across voice and SMS.上个月，Twilio 推出了 Agent-to-Human (A2H) 协议，旨在通过规范 AI 代理如何通过语音和短信与人类交互，来补充 MCP 和 Agent-to-Agent (A2A) 架构。

Again, we are not arguing that agentic voice startups can’t compete with legacy players. Instead, we see these companies employing moats stemming from the local number inventory necessary for agents to speak to humans. On top of that – in Twilio’s case – voice startups don’t have the customer context to make this a customer-friendly, market viable service.我们再次强调，我们并非认为语音助手初创公司无法与传统企业竞争。相反，我们认为这些公司利用了本地号码资源构建的竞争壁垒，这些号码资源是客服人员与真人沟通的必要条件。此外，以 Twilio 为例，语音助手初创公司缺乏必要的客户背景信息，因此难以打造出用户友好且具有市场竞争力的服务。

Signed, Sealed, Delivered已签署、已盖章、已交付

Contrary to popular opinion, we understand that software company “moats” often do not come from the sophistication of their technology, but rather from operational, regulatory, and compliance advantages.与普遍的看法相反，我们了解到软件公司的“护城河”往往不是来自其技术的复杂性，而是来自运营、监管和合规方面的优势。

In this vein, we think DocuSign (DOCU US) stands as a valuable member of the Agentic Ecosystem, given data and regulatory advantages. The company has embarked on a strategic shift from pure, seat-based E-Sign subscriptions to their Intelligent Agreement Management (IAM) platform.基于此，我们认为 DocuSign（DOCU US） 凭借其数据和监管优势，是代理生态系统中不可或缺的一员。该公司已开始战略转型，从纯粹的基于席位的电子签名订阅模式转向其智能协议管理（IAM） 平台。

The IAM platform was created to provide enterprises with AI tools that manage the entire contract lifecycle: Document creation, review, negotiation, signature, and analysis can be managed all through DocuSign’s IAM platform. Notably, this offering generated north of $350 million in annual recurring revenue for fiscal 2026, comprising roughly 11% of ARR. The company expects this segment to nearly double in fiscal 2027, reaching a high-teens share of total ARR.DocuSign 的 IAM 平台旨在为企业提供人工智能工具，以管理整个合同生命周期：文档创建、审核、协商、签署和分析均可通过 DocuSign 的 IAM 平台进行管理。值得注意的是，该产品在 2026 财年创造了超过 3.5 亿美元的年度经常性收入 (ARR)，约占 ARR 的 11%。公司预计该业务板块在 2027 财年将增长近一倍，达到 ARR 的 15% 以上。

The IAM product has also garnered partnership interest from agentic vendors such as Anthropic and OpenAI, who have integrated IAM through the model context protocol (MCP) or, in Anthropic’s instance, the Claude Skills connector.IAM 产品也引起了 Anthropic 和 OpenAI 等智能体供应商的合作兴趣，他们通过模型上下文协议 (MCP) 集成了 IAM，或者在 Anthropic 的例子中，通过 Claude Skills 连接器集成了 IAM。

“Last month, we partnered directly with Anthropic to make IAM available as part of Claude Cowork. The DocuSign MCP connector is available in beta today through Anthropic’s Connectors Directory. It enables DocuSign customers to use Cowork’s natural language prompts to automate agreement workflows and securely create, review, send and manage agreements in IAM, all with DocuSign’s trusted security and access controls. In addition to Cowork, IAM also connects via MCP server to OpenAI’s ChatGPT, Google Gemini, GitHub Copilot Studio and Salesforce’s Agentforce.”“上个月，我们与 Anthropic 直接合作，将 IAM 作为 Claude Cowork 的一部分推出 。DocuSign MCP 连接器现已在 Anthropic 的连接器目录中提供测试版。它使 DocuSign 客户能够使用 Cowork 的自然语言提示来自动化协议工作流程，并在 IAM 中安全地创建、审核、发送和管理协议，所有操作均在 DocuSign 可信赖的安全性和访问控制的保护下进行。除了 Cowork，IAM 还通过 MCP 服务器连接到 OpenAI 的 ChatGPT、Google Gemini、GitHub Copilot Studio 和 Salesforce 的 Agentforce。”

– DocuSign Q4 FY2026 Earnings Call– DocuSign 2026 财年第四季度财报电话会议

DocuSign’s advantage stems not just from over 200 million signed contracts and agreements, but also from its regulatory moat. DocuSign is both FedRAMP and GovRAMP certified, earning them valuable contract awards from bodies such as the US Department of War. What’s more, agentic workflows on DocuSign’s platform are underpinned by the US ESIGN Act, the Uniform Electronic Transactions Act (UETA), and Europe’s eIDAS, which permit digital signatures through DocuSign’s platform.DocuSign 的优势不仅源于其超过 2 亿份已签署的合同和协议，更在于其强大的监管壁垒。DocuSign 同时拥有 FedRAMP 和 GovRAMP 认证，因此赢得了包括美国战争部在内的众多机构的宝贵合同。此外，DocuSign 平台上的代理工作流程以美国《电子签名法案》(ESIGN Act ) 、《 统一电子交易法案》(UETA) 和欧洲《电子身份认证系统》( eIDAS )为基础，这些法案允许通过 DocuSign 平台进行数字签名。

3) Governance3）治理

All agents create data and expend resources, widening the surface area of the enterprise. As agentic adoption takes hold, however, the question in the boardroom shifts from “can they do it?” to “should they be doing it?”and “who is watching?”所有代理都会产生数据并消耗资源，从而扩大企业的覆盖范围。然而，随着代理技术的普及，董事会讨论的问题也从 “他们能做到吗？” 转变为 “他们应该这样做吗？” 以及 “谁在监督？”

An unharnessed agent can be outrageously expensive – both in what it consumes and what it exposes. An agent running 24/7 could potentially vacuum up a month’s worth of AWS spend in a single week, creating unexpected costs and distractions. Similarly, a rogue prompt injection can direct an agent to leak data, poison records, and abuse integrated tools. The ramifications of these attacks, if left unguarded, could cost an organization millions.不受控制的代理程序可能会造成极其高昂的成本——无论是在资源消耗还是安全风险方面。一个全天候运行的代理程序可能在一周内就消耗掉相当于一个月 AWS 的费用，造成意想不到的成本和干扰。同样，恶意注入指令可以指示代理程序泄露数据、篡改记录并滥用集成工具。如果不加以防范，这些攻击的后果可能给组织造成数百万美元的损失。

Volume without guardrails is a catastrophe waiting to happen. That’s where Governance becomes the focus. What was historically a bifurcated market between Observability vendors (what is it doing?) and Securityvendors (Is it allowed to do that?) has begun to converge into a single discipline – which we’ve coined Agentic Governance.缺乏监管的规模化应用注定会酿成灾难。正因如此，治理才显得尤为重要。过去， 可观测性供应商（ 它究竟在做什么？）和安全供应商（ 这样做是否合法？）之间存在着明显的两极分化，而如今，两者正逐渐融合为一个单一的领域——我们称之为 “智能体治理”。

The same telemetry that tells you an agent is underperforming is the same telemetry that tells you it’s been compromised. The identity function that authorizes agent access is the same function that will revoke access when behavior appears anomalous. This convergence is the most important structural shift in the infrastructure software market today.用于检测代理性能不佳的遥测数据，与用于检测代理是否已被入侵的遥测数据相同。用于授权代理访问的身份验证功能，也与用于在代理行为异常时撤销其访问权限的功能相同。 这种融合是当今基础设施软件市场最重要的结构性转变 。

Observability: Through the Looking Glass可观测性：透过镜子

Observability (collecting and analyzing logs, metrics, and traces) is more than just a monitoring and troubleshooting tool in the agentic world. As agents become more autonomous and ubiquitous, observability becomes the financial plumbing of modern architectures and an operational harness for a fresh breed of business models.在智能体领域，可观测性（收集和分析日志、指标和追踪数据）不仅仅是一种监控和故障排除工具。随着智能体变得更加自主和无处不在，可观测性将成为现代架构的财务基础，以及新型商业模式的运营工具。

In traditional software, you might purchase licenses and pay a fixed fee. But in an era where AI agents spin up and down on demand, if you can’t measure precisely who’s using what and when, you’ll never get an accurate handle on costs or revenue.在传统软件时代，您可能需要购买许可证并支付固定费用。但在人工智能代理按需启动和停止的时代，如果您无法精确衡量谁在何时使用什么，您将永远无法准确掌握成本或收入。

Ironically, turning on more detailed observability often raises your bills: just like installing a high-precision meter can reveal usage you didn’t know existed. Platforms like Datadog (DDOG US), New Relic, or Prometheus can generate significant data volume, which directly translates to higher cloud bills.具有讽刺意味的是，启用更详细的可观测性功能往往会导致账单上涨：就像安装高精度电表会揭示出你之前不知道的用电量一样。Datadog(DDOG US)、New Relic 或 Prometheus 等平台会产生大量数据，这直接导致云账单增加。

Yet these costs typically pale in comparison to what you can save (or earn) by catching inefficiencies early, billing customers for actual usage, or preventing unplanned downtime. The ROI is unlocked the moment you spot a costly flaw – or the moment you can precisely meter a usage-based service.然而，与及早发现低效环节、按实际使用量向客户收费或预防计划外停机所节省（或赚取）的费用相比，这些成本通常微不足道。 一旦发现代价高昂的缺陷，或者能够精确计量按使用量计费的服务，投资回报率便会立即显现。

Observability platforms will increasingly help product and development teams discover how models and agents are actually being used in the wild. Despite advances made in AI development over the last several years, one still doesn’t know how a model and its users will behave until it is deployed. That is, of course, only if behaviors, usage, and outcomes are observed and analyzed.可观测性平台将日益帮助产品和开发团队了解模型和智能体在实际应用中的使用情况。尽管人工智能开发在过去几年取得了长足进步，但在部署之前，我们仍然无法预知模型及其用户的行为。当然，前提是必须观察和分析其行为、使用情况和结果。

Security in a Zero Trust World零信任世界中的安全

Historically, cybersecurity was built around a stable set of assumptions: humans use devices, devices connect to networks, and the sprawl of devices dictates the “perimeter” of the organization. This was known as “Castle and Moat” architecture – once users were inside the perimeter, they were granted a high degree of implicit trust. The surface area of the enterprise was large, but it was fairly intuitive to protect and monitor.从历史上看，网络安全是建立在一套稳定的假设之上的：人使用设备，设备连接到网络，而设备的蔓延决定了组织的“边界”。这被称为 “城堡与护城河” 架构——一旦用户进入边界之内，他们就被赋予了高度的隐性信任。企业的安全防护面虽然很大，但保护和监控起来却相当直观。

The first crack in this model came with the explosion of cloud workloads and remote workers. Employees logged on from home networks, coffee shops, and personal devices. Meanwhile, data migrated from on-prem servers to SaaS applications built upon AWS, Azure, and Google Cloud. This shift ushered in a class of “next-generation” firewall companies built for the hybrid era. These businesses were able to survey traffic at the application layer, not just at the port level. This architecture was sufficient, although a fundamental loophole still remained: Once you were inside the firewall, you were trusted.这种模式的第一个裂痕出现在云工作负载和远程办公人员激增的时代。员工们通过家庭网络、咖啡馆和个人设备登录工作。与此同时，数据也从本地服务器迁移到基于 AWS、Azure 和 Google Cloud 构建的 SaaS 应用。这种转变催生了一批专为混合云时代打造的“下一代”防火墙公司。这些公司能够监控应用层流量，而不仅仅是端口层。这种架构虽然足够强大，但仍然存在一个根本性的漏洞：一旦进入防火墙内部，你就被信任了。

Zero Trust emerged as the successor architecture to the hybrid era. Rather than implicit trust based upon network access, Zero Trust operates under a “Never Trust, Always Verify” ethos. Every request from a human, device, application – must be continuously verified. The essence of Zero Trust boils down to: “Who are you?” and “Are you authorized to access this service?”零信任架构作为混合架构时代的后继者应运而生。它摒弃了基于网络访问权限的隐式信任，而是秉持“永不信任，始终验证”的理念。来自人、设备或应用程序的每一个请求都必须持续验证。零信任的核心在于：你是谁？ 你是否被授权访问此服务？

This turns out to be structurally compatible for the Agentic Era. Every meaningful agent inside an enterprise will need something resembling an employee file: who created it, what systems it can access, what secrets it is authorized to hold, which tools it can invoke, how it gets suspended, and what its chain of action looked like after an incident.事实证明，这种结构与智能体时代相契合。企业内部每个有意义的智能体都需要类似员工档案的东西：谁创建了它，它可以访问哪些系统，它被授权持有哪些机密信息，它可以调用哪些工具，它如何被停职，以及在发生事件后它的行动链是怎样的。

We have been long-standing Okta (OKTA US) bears – since October 2023, we’ve maintained that long NET / short OKTA would be long term winning trade (and it has been). However, Okta is already extending identity controls to non-human identities. It is “cheap”, and could see some rerating off of this narrative – although long term upside will be an execution issue and the company has not executed all that well. We’re keeping an eye on it, but we’re not convinced yet.我们长期以来一直看空 Okta（OKTA US）—— 自 2023 年 10 月以来 ，我们一直坚持做多 NET/做空 OKTA 才是长期盈利策略（事实也证明如此）。然而，Okta 已经开始将身份控制扩展到非人类身份。它目前估值“偏低”，并且可能基于这一发展趋势获得一些重新估值——尽管长期上涨空间取决于执行力，而该公司目前的执行力并不尽如人意。我们会继续关注，但目前仍未完全确信。

CyberArk (now PANW) has rolled out agent-specific privilege controls. Cisco is building AI BOM (bill of materials), MCP cataloging, and runtime protections around agentic tool use. Agentic identity management will become increasingly important.CyberArk（现更名为 PANW）已推出针对特定代理的权限控制。思科正在围绕代理工具的使用构建 AI 物料清单（BOM）、MCP 目录和运行时保护。代理身份管理将变得越来越重要。

A compromised agent doesn’t need to “break in”. Unlike a phishing attack – in which a cybercriminal spoofs an email or text message with the objective of gaining credentials – the agent is already within the bounds of the perimeter. All of the workflows undertaken by the agent are deemed to be normal unless someone, or something, flags it as anomalous.被入侵的代理无需“入侵”。与网络钓鱼攻击（网络犯罪分子伪造电子邮件或短信以获取凭证）不同，代理本身已处于安全边界内。除非有人或某种机制将其标记为异常，否则代理执行的所有工作流程都被视为正常。

So who operates the “something”? The answer is taking shape through a combination of platformization and M&A, as incumbent cybersecurity businesses aim to reinforce their product suite with a comprehensive toolbox of security capabilities.那么，谁在运营“这个东西”呢？答案正在通过平台化和并购的结合而逐渐形成，因为现有的网络安全企业旨在通过全面的安全功能工具箱来加强其产品组合。

Cloud-security giant Palo Alto Networks (PANW US) stands to benefit from this structural shift. Management’s recent acquisition spree clearly outlines where they see threats clustering in the Agentic Era.云安全巨头 Palo Alto Networks (PANW US) 有望从这一结构性转变中获益。管理层近期的一系列收购行动清晰地表明了他们对智能体时代威胁集中爆发点的看法。

Chronosphere, the more recent of the two pick-ups, expands the company’s reach into observability. The differentiation of the combined platform lies in handling high-cardinality data sets, framework-ready metrics, distributed tracing, and cost-aware telemetry pipelines that let customers filter and route data before observability spending gets out of hand. In an agentic environment, where workloads appear and disappear continuously, that is not just useful for debugging but a fundamental layer of cost control, revenue attribution, and automated response.Chronosphere 是两家公司近期收购的子公司之一，它拓展了公司在可观测性领域的业务范围。合并后的平台优势在于能够处理高基数数据集、框架就绪的指标、分布式追踪以及成本感知型遥测管道，使客户能够在可观测性支出失控之前过滤和路由数据。在工作负载不断出现和消失的代理环境中，这不仅有助于调试，而且是成本控制、收入归因和自动响应的基础层。

Zscaler (ZS US) is another pioneer of the Zero Trust architecture, having blended secure web gateways, data loss protection (DLP), and cloud access security into a single, inline proxy that brokers every connection between user and application – a posture known as SASE, or secure access service edge. The company’s Zero Trust Exchange (ZTE) acts as a virtual security escort for agentic traffic – ensuring that an agent is authorized to access a particular resource, at a particular time, for a particular reason.Zscaler（ZS US）是零信任架构的另一先驱，它将安全 Web 网关、数据丢失防护 (DLP) 和云访问安全融合到一个单一的内联代理中，该代理负责用户和应用程序之间的所有连接——这种架构被称为 SASE，即安全访问服务边缘。该公司的零信任交换 (ZTE) 充当代理流量的虚拟安全护卫，确保代理在特定时间出于特定原因获得授权才能访问特定资源。

If an agent is making API calls to external services, apps, or circulating data between tools – Zscaler’s inline security tool is the real time enforcement layer of this sequence.如果代理向外部服务、应用程序发出 API 调用，或在工具之间传递数据，则 Zscaler 的内联安全工具是此序列的实时执行层。

Conclusion结论

It may be easy to read this and say, “that’ll be cool… in 2030”. But the Agentic Era is not a newly contrived sci-fi meme. This sort of agentic explosion has long been assumed to be the natural evolution of AI in action, and one that has more far-reaching and disruptive implications than the chatbot era. Now, it’s not hypothetical – it’s happening.读到这里，你或许会想：“这要是到了 2030 年才出现，那可就酷毙了。” 但智能体时代并非什么新奇的科幻概念。这种智能体的爆发式增长，长期以来都被认为是人工智能实际应用的自然演进，而且其影响远比聊天机器人时代更为深远，更具颠覆性。如今，这不再是假设，而是正在发生。

At the same time, a number of yet-to-be-considered mines remain laid from threats to existing business models from artificial intelligence. While it’s simple to throw together a basket filled with software names and say “well, these probably will be used by AI”, we’ve attempted to navigate that minefield by narrowing this down to names that have limited risk and a clear angle for upside to this trend.与此同时，人工智能对现有商业模式的威胁仍存在诸多尚未考虑的潜在风险。虽然简单地罗列一堆软件名称并断言“这些软件很可能会被人工智能使用”看似简单，但我们已尽力避开这些雷区，将范围缩小到风险有限且具有明显增长潜力的软件。

Judging by the explosive uptake and usage of agentic platforms in just the past several months, we expect that implications will come faster than most expect – the train has left the station. Merely putting “agentic” on a slide deck won’t mean much, but the companies that truly enable this structural shift should see meaningful long-term growth. Our basket below captures broad thematic exposure to these Agentic Utilities.从过去几个月智能代理平台爆炸式增长和使用情况来看，我们预计其影响将比大多数人预想的来得更快——时代已经来临。仅仅在演示文稿中提及“智能代理”意义不大，但真正推动这一结构性转变的公司应该会获得显著的长期增长。我们下面的投资组合涵盖了这些智能代理实用工具的广泛主题。

The Agentic Utilities Basket代理实用工具篮

This article is for informational purposes only and does not constitute investment advice. By accessing this material, you agree to our Terms of Service.本文仅供参考，不构成投资建议。访问本文即表示您同意我们的服务条款 。

更多有价值内容详见社群

社群到底提供什么？内容来自哪里？这篇说清楚