2026大构想：代理式交互界面

我最近与IT部门负责人聊过，他告诉我，在长达二十年的职业生涯中，他第一次认为IT支持将发生根本性变革。

I chatted with the head of IT recently who told me for the first time in his two decade long career, he believed that IT support was fundamentally gonna change.

如果我们所有人都希望这款软件为我们工作，理想情况下，它的能力至少应与人类相当，甚至更强。

If all of us want this software to be doing work for us, ideally, it's doing work with at least, if not, more competency than a human could.

我们不再为人类设计，而是为代理设计。

We're no longer designing for humans, but for agents.

新的优化不再是视觉层次，而是机器可读性。

The new optimization isn't visual hierarchy, but machine legibility.

这将改变我们创造的方式以及我们所使用的工具。

And that will change the way we create and the tools that we use to do it.

每年，我们都会退后一步，提出一个简单的问题。

Every year, we step back and ask a simple question.

接下来，构建者会关注什么？

What will builders focus on next?

我们的2026年重大构想汇集了投资团队认为将塑造来年科技趋势的主题。

Our twenty twenty six big ideas bring together the themes our investing teams believe will shape the coming year in tech.

本集围绕三个核心理念展开，它们共同揭示了AI产品下一步的真实发展方向。

This episode is built around three big ideas that, together, explain where AI products are actually heading next.

这种转变不仅仅是模型变得更聪明了，更是软件形态正在发生变化。

The shift is not just that models are getting smarter, it's that software is changing shape.

AI正从你咨询的工具，转变为能够理解意图并采取行动的系统。

AI is moving from a tool you consult to a system that can understand intent and take action.

你将听到三种不同的视角，探讨这一转变对界面意味着什么、对软件和信息设计意味着什么，以及对组织内部工作执行方式意味着什么。

You're going to hear three different perspectives on that transition: what it means for the interface, what it means for how we design software and information, and what it means for how work gets executed inside organizations.

第一个核心理念是，提示框并非AI的最终界面。

The first big idea is that the prompt box is not the final interface for AI.

马克·安德鲁斯科认为，获胜的产品将不再像聊天机器人，而更像主动的团队伙伴。

Marc Andrusko argues that the winning products will feel less like chat and more like proactive teammates.

它们会注意到你在做什么，预判你的需求，并提出你可以批准的操作。

They'll notice what you're doing, anticipate what you need, and propose actions you can approve.

以下是马克。

Here's Marc.

我是马克·安德鲁斯科，我们AI应用投资团队的合伙人。

I'm Mark Andrusko, a partner on our AI apps investing team.

我对2026年的核心观点是：提示框将不再是AI应用的主要用户界面。

My big idea for 2026 is the death of the prompt box as the primary user interface for AI applications.

下一代应用将需要少得多的提示。

The next wave of apps will require way less prompting.

它们会观察你的行为，并主动提供你需要审查的操作建议。

They'll observe what you're doing and intervene proactively with actions for you to review.

我们所瞄准的机遇，过去是全球每年3000亿至4000亿美元的软件支出。

The opportunity we're attacking used to be the 300 to $400,000,000,000 of software spend annually in the world.

现在让我们兴奋的是，仅在美国就存在的13万亿美元的劳动力支出。

Now what we're excited about is the $13,000,000,000,000 of labor spend that exists in The US alone.

这使得软件的市场机会或总潜在市场规模扩大了约30倍。

It's made the market opportunity or the TAM for software about 30 times bigger.

如果你从这一点出发，再进一步思考：如果我们所有人都希望这些软件为我们工作，理想情况下，它至少应该具备，甚至超越人类的工作能力。

If you start from there and then you think about, okay, if all of us want this software to be doing work for us, ideally, it's doing work with at least if not more competency than a human could.

对吧？

Right?

所以我喜欢思考，最好的员工是怎么做的？

And so I like to think about like, well, what do the best employees do?

最好的人类员工是怎么做的？

What do the best human employees do?

我最近一直在谈论一个在推特上流传的图表。

And I've recently been talking about this graphic that was floating around on Twitter.

它是一个金字塔，展示了五种类型的员工，以及那些最具自主性的员工为何最优秀。

It's a pyramid of, like, the five types of employees and the ones with the most agency and why they're the best.

如果你从金字塔的底层开始，那是那些发现问题后来找你、询问该怎么办的人。

So if you start at the bottom rung of the pyramid, it's like people who identify a problem and then come to you and ask for help and ask what to do.

这是自主性最低的员工。

And that's like the lowest agency employee.

但如果你往上走到S级，也就是你可能拥有的最具自主性的员工，他们会先发现问题。

But if you go to the s tier, like the most high agency employee you could possibly have, they identify a problem.

他们会进行必要的研究，以诊断问题的根源。

They do research necessary to diagnose where the problem came from.

他们会探索多种可能的解决方案。

They look into a number of possible solutions.

他们会实施其中一种方案，并随时向你汇报进展。

They implement one of those solutions, and then they keep you in the loop.

或者他们会在最后一刻才来找你，问：‘你同意我找到的这个方案吗？’

Or they come to you at the very last minute and say, like, do you approve this solution I found?

这就是我认为AI应用的未来会是的样子。

And that's what I think the future of AI apps will be.

我认为这也是每个人所期望的。

And I think that's what everyone wants.

这是我们所有人都在努力追求的目标。

That's what we're all working towards.

因此，我很有信心，随着大语言模型持续变得更好、更快、更便宜，我认为在高风险场景中，用户行为仍会要求在最后阶段保留人类参与，以最终批准各项决策。

So I feel pretty confident that we're almost I think LLMs have continued to get better and faster and cheaper, and I think there's a world in which the user behavior will still necessitate a human in the loop at the very end to sort of approve things certainly in high stakes contexts.

但我认为，这些模型已经完全有能力达到这样的程度：它们能代表你提出非常聪明的建议，而你只需点击确认即可。

But I think the models are more than capable of getting to a point where it's suggesting something really smart on your behalf, and you basically just have to click accept.

正如你们所知，我非常痴迷于‘AI原生CRM’这个概念，我认为这正是这类主动型应用的理想范例。

As you guys know, I'm pretty obsessed with the notion of an AI native CRM, and I think this is, like, a perfect example of what these proactive applications could look like.

在当今的世界里，销售人员可能会打开他们的CRM系统，浏览所有未完成的商机，查看当天的日程，并思考：现在我能采取哪些行动，以最大程度地影响我的销售漏斗和成交能力？

So in today's universe, a salesperson might go open their CRM, explore all the open opportunities they have, look at their calendar for that day, and try to think about, okay, what are the actions I can take right now to have the greatest impact on my funnel and my ability to close deals?

而在未来的CRM系统中，你的AI代理或AI-CRM应能持续不断地代你完成所有这些工作，不仅识别出你销售线索中显而易见的机会，还会翻阅你过去两年的所有邮件，发现曾经是潜在客户但被你忽视的线索。

With the CRM of tomorrow, your AI agent or your AI CRM should be doing all these things on your behalf in perpetuity, identifying not only like the most obvious opportunities that are in your pipeline, but going through your emails from the last two years and harvesting, you know, this was once a warm lead and you kinda let it die.

比如，我们是不是该给他们发一封邮件，重新激活他们，让他们回到你的流程中？

Like, maybe we should send them this email to drum them back up into your process.

对吧？

Right?

我认为，撰写邮件、整理日程、查阅过往通话记录，这类应用场景的可能性几乎是无穷无尽的。

So I think there are so many ways in which drafting an email, harvesting your calendar, going through your old call notes, the opportunities are just endless.

普通用户在绝大多数情况下，仍然希望保留最后一道审批权。

The ordinary user will still want that last mile approval almost a 100% of the time.

他们希望人类在环中的部分仍然是最终决策者。

They will want the human part of the human in the loop to be the final decision maker.

这很好。

And that's great.

我认为这正是这一技术自然演进的方式。

I think that's, like, the natural way in which this will evolve.

我可以想象这样一个世界：高级用户会付出大量额外努力，训练他们所使用的AI应用，使其尽可能全面地了解他们的行为和工作方式。

I could imagine a world in which the power user is basically taking a lot of extra effort to train whichever AI app it's using to have as much context about their behavior and how they perform their work as humanly possible.

这些应用将利用更大的上下文窗口。

These will utilize larger context windows.

这些应用将利用嵌入在众多大语言模型中的记忆功能，使高级用户能够真正信任应用完成99.9%的工作，甚至100%，他们会以无需人工审批即可完成的任务数量为傲。

These will utilize memory that's been baked into a lot of these LLMs and make it such that the power user can really trust the application to do 99.9% of the work or maybe even a 100, and they'll pride themselves on the number of tasks that get done without a human needing to approve them.

马克指出了界面从提示到执行的转变。

Mark gives the interface shift from prompting to execution.

第二个重要理念自然随之而来。

The second big idea follows naturally.

如果代理代表我们导航软件，那么我们就必须开始设计软件，使其能够被代理理解。

If agents are the ones navigating software on our behalf, then we have to start building software to be understood by them.

斯蒂芬妮·张称这种软件为机器可读软件。

Stephanie Zhang calls this machine legible software.

在以代理为先的世界里，视觉层次的重要性降低，而结构的重要性提升。

In an agent first world, visual hierarchy matters less and structure matters more.

优势将转向那些机器能够可靠解析和操作的产品、内容和系统。

The advantage shifts to products, content, and systems that machines can reliably interpret and operate inside.

以下是斯蒂芬妮。

Here's Stephanie.

你好。

Hi.

我叫斯蒂芬妮·邓，我是16z增长团队的投资合伙人。

My name is Stephanie Dang, and I'm an investing partner on the a sixteen z growth team.

我对2026年的核心观点是：为代理而设计，而非为人类。

My big idea for 2026 is creating for agents, not for humans.

我对2026年最期待的一件事是，人们必须开始改变他们的创作方式。

Something I'm super excited about for 2026 is that people have to start changing the way they create.

这涵盖了从创作内容到设计应用程序的方方面面。

And this ranges from creating content to designing applications.

人们开始通过代理作为中介来与网络或应用程序等系统进行交互。

People are starting to interface with systems like the web or their applications with agents as an intermediary.

对人类消费重要的东西，对代理消费而言将不再以相同的方式重要。

And what mattered for human consumption won't matter the same way for agent consumption.

我上高中时学过新闻学。

When I was in high school, I took journalism.

在新闻学中，我们学到在新闻文章的导语中要先交代五个W和一个H，并在特稿中以吸引人的开头开场。

And in journalism, we learned the importance of starting with the five Ws and Hs in the lead paragraph for news articles and to start with a hook for features.

为什么？

Why?

为了吸引人类的注意力。

For human attention.

也许人类会忽略藏在第五页的深刻相关且富有洞察力的陈述，但代理不会。

Maybe a human would miss the deeply relevant, insightful statement buried on page five, but an agent won't.

多年来，我们一直针对可预测的人类行为进行优化。

For years, we've optimized for predictable human behavior.

你希望成为谷歌搜索结果中的前几条。

You want to be one of the first search results back from Google.

你希望成为亚马逊列表中的前几项。

You want to be one of the first items listed on Amazon.

这种优化不仅适用于网页，也适用于我们设计软件时。

And this optimization is not just for the web, but as we design software too.

应用程序是为人类的视觉和点击而设计的。

Apps were designed for human eyes and clicks.

设计师们优化了良好的用户界面和直观的流程。

Designers optimized for good UI and intuitive flows.

但随着代理使用的增加，视觉设计在整体理解中的重要性逐渐降低。

But as agent usage grows, visual design becomes less central to overall comprehension.

以前，在发生故障时，工程师们会进入他们的Grafana仪表板，试图拼凑出发生了什么。

Before, during incidents, engineers would go into their Grafana dashboards and try to piece together what was going on.

现在，AI SRE会接收遥测数据，分析这些数据，并直接在Slack中向人类报告假设和洞察。

Now AI SREs take in telemetry data, they'll analyze that data, and they'll report back with hypotheses and insights directly into Slack for humans to read.

以前，销售团队必须逐点点击并导航Salesforce或其他CRM系统来获取信息。

Before, sales teams would have to click through and navigate Salesforce or other CRMs to gather information.

现在，代理会提取这些数据并为他们总结出洞察。

Now agents will take that data and summarize insights for them.

我们不再为人类设计，而是为代理设计。

We're no longer designing for humans, but for agents.

新的优化不再是视觉层次，而是机器可读性。

The new optimization isn't visual hierarchy, but machine legibility.

这将改变我们创造的方式以及我们用来实现它的工具。

And that will change the way we create and the tools that we use to do it.

我们不知道代理究竟在寻找什么，但我们知道，代理在阅读文章全文方面远比人类更出色，而人类可能只读前几段。

It is a question we don't know the answer to what agents are looking for, but all we know is that agents do a much better job at, you know, reading all of the text in an article versus maybe a human would just read, you know, the first couple paragraphs.

现在有很多工具，不同组织使用这些工具来确保当消费者在向ChatGPT询问最佳企业信用卡或最佳鞋子时，他们能够出现。

There are a bunch of tools out there that different organizations use to just make sure that they show up when consumers are prompting ChatGPT asking for the best corporate card or the best shoes to buy.

因此，市场上存在大量我们称之为地理工具的工具，人们正在使用它们。

And so there's like a bunch of what we call geo tools out there in the market that people are using.

但每个人都想知道：AI代理希望看到什么。

But everybody is asking the question what AI agents want to see.

我喜欢这个问题。

I love this question.

当人类可能完全退出循环时，我们已经看到在某些情况下这种情况正在发生。

When humans may choose to exit the loop entirely, we're already seeing that happen in some cases.

我们的投资组合公司Decagon已经能够自主回答许多客户的问题。

Our portfolio company Decagon is answering questions for a lot of their customers already autonomously.

但在其他情况下，比如安全运营或事件解决，我们通常看到更多的人工参与：AI代理首先尝试找出问题所在，进行分析，并向人类提供不同的潜在情况。

But for other cases, security operations or incident resolution, we typically see a little bit more human in the loop where the AI agent takes first stab at trying to figure out what the issue is, running the analysis and serving to the humans different potential situations.

这些通常是责任更高、分析更复杂的案例，我们看到人类仍会留在循环中，并且在模型和技术达到极高的准确率之前，人类很可能会长期保持参与。

Those tend to be cases of higher liability, more complex analyses that we see humans staying in the loop and will probably stay in the loop for much longer until the models and the technology get to incredibly high accuracy.

我不知道代理是否会观看Instagram短视频。

I don't know if agents will be watching Instagram Reels.

这真的很有趣。

It's really interesting.

至少在技术层面，优化机器可读性、洞察力和相关性至关重要，而过去更多是通过炫目的方式吸引人们的注意力。

At least on the tech side, it is really important to optimize for that machine legibility piece, optimize for insight, optimize for relevance, especially versus in the past, it was more about hooking people in, capturing attention in flashy ways.

我们已经看到的是高容量、高度个性化的内容。

What we're seeing already is case of high volume, hyper personalized content.

也许你不是创作一篇极其相关、极具洞察力的文章，而是大量生产低质量的内容，但覆盖了你认为代理可能想看到的各种话题，这几乎相当于过去的关键词。

And maybe you don't create one extremely relevant article, extremely relevant and insightful article, but maybe you're creating extremely high volumes of low quality content, but addressing different things that you may think an agent wants to see, almost like the equivalent of keywords.

在代理时代，内容创作的成本几乎降为零，轻松就能生成大量内容，这可能导致一个潜在风险：通过海量内容来争夺代理的注意力。

In the era of agents where cost of creation of content kind of goes to zero and it's really easy to create high volumes of content, that's a potential risk around just high volumes of things to be able to try to capture agent attention.

如果软件变得机器可读，且代理能够在不同工具间执行任务，那么最大的挑战就不再是表面的。

If software becomes machine legible and agents can execute tasks across tools, then the biggest challenge is not cosmetic.

而是组织层面的。

It's organizational.

这引出了第三个重要观点。

That leads to the third big idea.

莎拉·王描述了代理层的兴起，该层位于传统记录系统之上，成为实际工作发生的地方。

Sarah Wang describes the rise of an agent layer that sits above the traditional system of record and becomes the place where work actually happens.

它缩短了意图与执行之间的距离，改变了软件系统控制流程的方式。

It collapses the distance between intent and execution in changes which software systems control the flow.

以下是莎拉。

Here's Sarah.

我是莎拉·王，a16z增长基金的普通合伙人，我对于2026年的核心观点是，记录系统将开始失去其优势。

I'm Sarah Wang, general partner on a16z growth, and my big idea for 2026 is that systems of record start to lose their edge.

当代理能够独立执行已签署的意图时，被动的记录层就不再有意义了。

A passive system of record layer stops making sense when agents can independently execute on an signed intent.

我预计会出现一个全新的动态代理层，它对员工来说比传统的记录系统更有意义。

I expect to see a new dynamic agent layer that actually makes sense for employees to replace legacy systems of record.

这是在将智能融入企业这一漫长道路上非常令人兴奋的进展。

This is a very exciting development on the long road of inserting intelligence into companies.

我并不是轻率地说系统记录正在失去其优势。

I don't say that systems of record are losing privacy lightly at all.

我曾经在一家公司工作，该公司几乎只投资于ERP和其他系统记录，因为数据引力具有很强的粘性。

I used to work at a firm that almost exclusively invested in ERPs and other systems of record because of the stickiness of the data gravity.

曾有一波SaaS 2.0公司获得大量融资，试图通过更优的用户界面挑战系统记录，但最终都失败了。

There was a wave of SaaS two point o that was well funded and tried and failed to take on the system of record, mostly through a better UI.

这是我们第一次看到真正威胁到系统记录的存在，原因在于意图与执行之间的距离正在消失。

This is the first time that we've seen a genuine threat to that, and that's because the distance between intent and execution is collapsing.

这并不是为用户带来20%到50%的体验提升，而是如何达到那种神奇的TEDx时刻。

And that's creating not a 20 to 50% better experience for the user, but how you get to that magical TEDx.

让我们以ITSM（IT服务管理）这一具体例子为例。

Let's take the concrete example of ITSM, IT service management.

这传统上一直是强大公司ServiceNow的领域。

This has traditionally been the domain of powerhouse company ServiceNow.

我最近与一位IT负责人交谈，他告诉我，在长达二十年的职业生涯中，他第一次相信IT支持将发生根本性变革。

I chatted with the head of IT recently who told me for the first time in his two decade long career, he believed that IT support was fundamentally gonna change.

五年后，它将完全不一样。

It will look completely different in five years.

那么为什么会这样呢？

So why is that?

如果你想想旧系统的工作方式，比如在公司里申请新软件权限需要多长时间，再对比一下正在出现的ITSM代理。

If you think about the way that the old systems work, how long it takes to do something like request access to new software in the firm, and you contrast that with the ITSM agents that are arriving.

它们接入你的技术栈后，这类请求几乎会瞬间完成。

They plug in your stack, and this type of request becomes nearly instantaneous.

通过大语言模型的进步，你现在可以提取意图。

Through advancements in LLMs, you can now extract intent.

你可以对请求类型进行分类。

You can classify the request type.

你可以将其映射到已知的工作流程，识别用户实体，从而使用户的请求以高效且准确的方式得到满足。

You can map it to a known workflow, identify user entities, and the request from the user becomes fulfilled in a way that is efficient and accurate.

因此，我们认为这种新范式中存在几个有价值的层面。

So we think there's a couple of valuable layers in this new paradigm.

当然，还有基础模型层。

Of course, there's the foundation model layer.

我们相信这一层仍然很有价值。

We believe that stays valuable.

但真正带来未来价值的是新兴的代理层，它尽可能贴近用户，收集并理解用户偏好。

But it's really the emerging agent layer that sits as close as possible to the user and is collecting data on that user, understanding user preferences that we think accrues value in the future.

根据我们所观察到的实际情况，我们认为这对新进入者来说是一个巨大的机遇。

Based on everything that we're seeing in the wild, we believe this is a huge opportunity for new players to come in and win.

为什么这么说？

Why is that?

我们现在正处于一个产品每周，甚至每天都在进步的阶段，你需要能够快速行动的团队。

We're in a phase right now where the product is getting better on a weekly, if not daily basis, and you need teams that move fast.

如果你要将意图与执行结合起来，那么连接这两者的，实际上是为客户提供准确可靠的解决方案。

If you're gonna collapse intent and execution, what bridges that is actually having an accurate, reliable solution for your customer.

否则，用户是不会使用的。

Otherwise, they're not gonna use it.

他们不会信任你所构建的代理。

They're not gonna trust the agent that you're building.

这就是为什么我们看到，即使是构建在 Datadog 等经典知名平台之上的代理，也输给了 Resolve 或 Traversal 等新兴 AI SRE 公司。

That's why we're starting to see even agents built on top of classic iconic platforms like Datadog lose to some of the new AI SRE companies like a Resolve or a Traversal.

我们对这一机遇感到非常兴奋，2026 年将是动态代理层超越系统记录的一年。

We're extremely excited about this opportunity, and 2026 is going to be the year that the dynamic agent layer overtakes the system of record.

综合来看，这三个核心理念构成了一个完整的故事。

Taken together, these three big ideas form a single story.

首先，界面从聊天转向行动。

First, the interface shifts from chat to action.

其次，设计从以人类为中心转向以代理可读为中心。

Second, the design shifts from human first to agent readable.

第三，工作流程从系统记录转向将意图转化为成果的代理层。

Third, the workflow shifts from systems of record to agent layers that turn intent into outcomes.

这就是这里‘代理式’的真正含义：AI 不再是你询问的东西，而成为主动执行的东西。

This is what agentic really means here: AI stops being something you ask and becomes something that does.

感谢收听本集的a16z播客。

Thanks for listening to this episode of the a 16 z podcast.

如果你喜欢本集，请务必点赞、评论、订阅、给我们打分或留下评价，并分享给你的朋友和家人。

If you like this episode, be sure to like, comment, subscribe, leave us a rating or a review, and share it with your friends and family.

如需收听更多集数，请前往YouTube、Apple Podcasts和Spotify。

For more episodes, go to YouTube, Apple Podcasts, and Spotify.

在X上关注我们@a16z，并在a16z.substack.com订阅我们的Substack。

Follow us on x at a sixteen z, and subscribe to our Substack at a16z.substack.com.

再次感谢收听，我们下集再见。

Thanks again for listening, and I'll see you in the next episode.

提醒一下，本内容仅作信息参考，不应被视为法律、商业、税务或投资建议，也不应用于评估任何投资或证券，且并非针对任何a16z基金的投资者或潜在投资者。

As a reminder, the content here is for informational purposes only, should not be taken as legal business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any a sixteen z fund.

请注意，a16z及其关联方可能仍持有本播客中讨论的公司的投资。

Please note that a sixteen z and its affiliates may also maintain investments in the companies discussed in this podcast.

如需更多详情，包括我们的投资链接，请访问a16z.com/disclosures。

For more details, including a link to our investments, please see a 16z.com forward slash disclosures.