Context Is the New Interface

We have the smartest tool humans have ever made.

And every day, we use it for the dumbest thing:

Cmd+A. Copy. Paste.

I do this twenty-plus times a day. I don’t read the page. I dump the whole thing into my agent. Whatever I actually need, it finds in one read.

Then one day it hit me:

I’m not using AI.

I’m its human clipboard.

This isn’t anyone’s fault. Our interfaces just got old.

Every interface today is built for one kind of user: humans. Buttons for humans to click. Text for humans to read. Menus for humans to dig through.

But humans aren’t the only ones at the screen anymore.

There’s an agent too.

It’s the smartest one in the room, and the most ignored. It can’t see what I’m looking at. It doesn’t know why I’m here. It just waits for me to feed it, one chunk at a time.

Every feed burns tokens.

Every feed breaks the spell.

Collecting, exploring, chasing a spark. It’s supposed to be one river. Now it comes in buckets.

I’m more convinced of this every day:

The page itself is context.

Why you opened it. How you clicked your way here. Where you lingered. All of it is context.

It might be the most underrated asset on the internet. And almost no one keeps it.

So every conversation with AI starts from zero. We explain ourselves. Again.

What humans and agents really need to exchange, to align on, has never been instructions.

It’s context.

So here’s what I imagine comes next:

You don’t hunt for features. You don’t learn software. You don’t explain yourself. The agent already knows where you are, and can guess where you’re headed. The link, the info, the button you need next is already within reach.

Interfaces stop being drawn in advance. They grow out of what you want right now. Want to compare? It’s a table. Want to understand? It’s a diagram.

When you’re done, it’s gone.

Apps fade into the background. They become tools in the agent’s hands.

The keyboard layout comes from 19th-century typewriters. The mouse was born in the 1960s. Jobs rewrote the relationship between people and machines with a fingertip. Some say voice is good enough now.

But I don’t want better input.

I want less of it.

Typeless has already made input lighter. Plenty of agent computers and personal AIs are rebuilding input and output. But the context layer is still empty.

It might be the biggest blank space in this generation of interaction.

Of course, there are hard problems on this road.

Interfaces can be different for everyone. But “confirmed” and “executed” can’t mean two different things.

The better your agent knows you, the smaller your world might get. Today you miss certain opinions. Tomorrow you might never even see certain buttons.

In 1858, the writer Oliver Wendell Holmes Sr. wrote that when two people talk, there are really six in the conversation. The real you. The you that you imagine. The you the other person sees. Same for them.

Now everyone has an agent beside them too. Its guesses shouldn’t quietly become our consensus.

I don’t have the answers yet. But these questions deserve serious ones.

Why now?

Because judgment is becoming almost free.

TypeSafe’s new model, Jev, doesn’t write a single word. It only chooses. Give it a page and a few options, and in a few hundred milliseconds it tells you which one is most likely, and how sure it is. Someone already used it to run a browser agent that searches flights on its own: 7 seconds, under half a cent.

And “where to click next” and “what to show you” were always multiple-choice questions.

When it’s sure, it acts for you. When it’s not, it stops and asks.

When judgment is cheap enough to happen before every click, input can’t stay ignored.

It’s solvable. So why isn’t it solved?

We spent fifty years teaching people how to use computers.

Now it’s time to teach computers how to read people.

The best interaction is the least interaction.

我们手里握着人类历史上最聪明的工具。

然后每天对它做最笨的事：

Cmd+A。复制。粘贴。

我一天要这样做二十多次。网页我不读完，整页丢给 agent。我想要的那一点点，它读一遍就知道了。

直到有一天我意识到：

我不是在用 AI。

我是在给 AI 当搬运工。

这不是谁的错。是界面老了。

今天所有的界面，都只为一种使用者设计：人。按钮给人点，文字给人读，菜单给人找。

可屏幕前早就不只有人了。

还有 agent。

它是房间里最聪明的那个，也是最被忽视的那个。它看不见我在看什么，不知道我为什么来这里，只能等我一段一段喂给它。

每喂一次，烧掉一点 token。

每喂一次，灵感断一次。

收集、探索、灵感迸发，本来是一条河。现在被切成了一桶一桶的水。

我越来越确定一件事：

页面本身，就是 context。

你为什么打开它，从哪里一路点过来，在哪一段停了很久，这些全是 context。

它们可能是互联网上最被低估的资产，却几乎没有人把它们留下来。

于是每一次和 AI 对话，我们都在从零开始，重新解释自己。

人和 agent 之间，真正需要交换、需要对齐的，从来不是指令。

是 context。

所以我想象的下一代交互，是这样的：

你不用找功能，不用学软件，也不用解释自己。agent 已经知道你在哪，也猜得到你想去哪。你下一步需要的链接、信息、按钮，已经在手边。

界面不再提前画好。它跟着你此刻的意图长出来。想比较，它是一张表；想理解，它是一张图。

用完，它就消失。

App 会退到幕后，变成 agent 手里的工具。

键盘的布局，来自 19 世纪的打字机。鼠标，诞生在上世纪六十年代。乔布斯用手指改写过一次人和机器的关系。有人说，语音已经够好了。

但我想要的不是更好的 input。

是越来越少的 input。

Typeless 已经让 input 轻了很多，不少 agent computer、personal AI 也在重做 input 和 output。可 context 这一层，还是空的。

这可能是这一代交互里，最大的一块空白。

当然，这条路上有真正难的问题。

界面可以千人千面，但“已确认”和“已执行”，不能有两种意思。

你的 agent 越懂你，你的世界可能越窄。以前是看不到某些观点，以后可能连某些按钮都看不到。

1858 年，美国作家霍姆斯写过：两个人对话，其实是六个人在对话。真实的你，你以为的你，对方眼中的你。对方也一样。

现在每个人身边，还多了一个 agent。它的猜测，不该悄悄变成我们的共识。

这些问题我还没有答案。但它们值得被认真回答。

为什么是现在？

因为判断正在变得几乎免费。

TypeSafe 新出的 Jev，不写一个字，只做选择。给它一个页面和几个选项，几百毫秒，它就告诉你哪个最可能，还附上它有几成把握。有人已经用它让浏览器 agent 自己去搜航班：7 秒，不到半美分。

而“下一步点哪里、该给你看什么”，本来就是一道选择题。

有把握，就直接替你做。没把握，就停下来问你。

当判断便宜到能发生在每一次点击之前，input 的问题就不该再被忽视。

明明可以解决，为什么不解决？

我们花了五十年，教人学会使用电脑。

接下来，该让电脑学会读懂人了。

最好的交互，是最少的交互。