商业产业 Hacker News AI 2026-10-04T22:38 正文 ok 1052 words

我们的 AI Agent 应该向帮助它们的人付费

英文标题:🔶 Our AI Agents Should Pay the People Who Help Them

作者认为,AI Agent 可以成为低摩擦的内容价值交换基础设施:创作者在内容旁发布付款请求,用户的数字助手则从预设月度预算中自动向有价值的内容提供者支付小额款项。作者以已于 2023 年关闭的 Flattr 为例,指出其按月度预算分配打赏的模式合理,但账户注册、网站接入和手动点击等摩擦阻碍了普及。HTTP 402、Coinbase x402 和 Cloudflare Pay Per Crawl 等机制已为机器间内容付费提供部分基础,但作者更关注个人助手使用个人预算,向具体帮助自己的创作者付款。对于开发者而言,这意味着 Agent 的支付能力、内容支付端点、支出上限和可审计日志,可能成为未来 Agent 基础设施的重要组成部分。
打开原文相关 3 · 机会 1
media222.8 KB media232.2 KB

中文

一幅炭笔画:一只巨大的紫色机器人手将一串串赭色小硬币倾泻给四位站在岩石平台上的人类创作者;他们身旁摆着各自的作品:画架上的素描、一叠手写稿纸、一本翻开的书和一把鲁特琴

你知道这一波 Agent 热潮中,最让我兴奋的是什么吗?作者能够直接在自己的内容旁发布付款请求,并附上一个用于提交小额支付的接口。

我想要这个已经很久了。2015 年,我在 Social Micropayments 中写过这个想法;2016 年,我又在另一篇文章中讨论过它。当时我描述了一种数字助手:它会观察你喜欢什么,并从月度预算中向创作者付款。我想,我对这件事的看法和当年基本一样。新的变化是,Agent——真正的 Agent,也就是我在自己的环境中整天运行的那种——如今真的已经出现,而且正在变得越来越好。

Flattr 的思路是对的

很久以前有一个名为 Flattr 的系统。它由 Peter Sunde(The Pirate Bay 创始人之一)和 Linus Olsson 于 2010 年创办,其运作方式是让用户预先注册一个 Flattr 账户——而这就是第一个问题。如果你访问的网站也接入了 Flattr,你就可以点击 Flattr 按钮;系统会从你的月度 Flattr 预算中划出一定比例,发送给对方。

因此,如果你的预算是 10 美元,而你当月对十篇文章进行了 Flattr,那么每位作者可以得到 1 美元。如果你只喜欢一项内容,那么那位作者将得到全部 10 美元。我很喜欢这个模式。我现在仍认为它的基本思路是正确的。Flattr 于 2023 年关闭。

问题在于摩擦,而且无处不在。你需要账户,创作者需要账户,网站需要部署按钮,而且你还得记得在阅读时实际去点击它。这些步骤中的每一步都会让用户流失,事实也确实如此。

让助手来做

我认为,我们可以实现类似的机制,但所有环节都由数字助手来协调。它会知道我们每个月愿意给别人多少钱,而我们则会拥有某种支付机制,可以是通过 Stripe 的小额支付,也可以是任何能够妥善处理这件事的平台。

任何触达该内容的主体,都可以选择通过小额支付为其付费;如果愿意,甚至可以支付更大金额。如果用户或 Agent 发现这份内容对自己或其委托人极其有用,它可以多付一点。

由于 Agent 全天候工作,并且能够同时处理如此多的事情,我的助手可以替我开展研究,并向为该研究提供最佳内容的人支付 2 美元、5 美元或 12 美元。而我不需要介入这一部分。老实说,我认为这才是最重要的部分,因为“记得去点击”一直都是整套机制最终失效的地方。

实现它需要什么

这个想法的妙处在于,几乎不需要太多人工步骤:

  • 创作者将自己的网站配置为支持支付端点:本质上是在内容旁放置一个定价请求,或“按你认为它值多少钱支付”的请求。

  • 用户将其运行 Agent 的 harness(无论具体是什么)配置为支持支付。

  • 用户与其 Agent 就愿意以这种方式花多少钱达成某种约定,包括每月限额和单项限额。

之后,它就会作为 Agent 活动的一部分自然发生。Agent 阅读某些内容,并将其用于为我完成某项任务;如果那篇内容最终发挥了重要作用,它就向撰写它的人付款。如果没有发挥作用,那么根据我们达成的约定,它可能支付几美分,也可能什么都不付。

坦率地说,我还不确定正确的支付通道应该是什么。它可能是 Stripe,可能是稳定币,也可能是某种尚不存在的东西。我也不知道创作者该如何定价,或者他们中的大多数人是否会干脆保留“按你认为它值多少钱支付”这一选项,让 Agent 自己决定。我有点怀疑最后一种方式会成为主流。

我还希望能够设定优先级,就像我在 2016 年所描述的那样。改变我思考问题方式的内容,应该比仅仅让我觉得好笑的内容获得更多预算。而预算就是护栏。一个能够花钱的 Agent 必须有严格的月度上限,并且要有我真正能看懂的日志;否则,它很快就会变成一个安全问题。

部分底层设施已经存在

当我在 X 上发布这个想法时,我问是否有人知道正在朝这个方向发展的系统,而现在确实已经有一些组件出现了。

HTTP 自 20 世纪 90 年代起就在规范中保留了 402 Payment Required 状态码。它被明确标注为“保留供未来使用”。Coinbase 于 2025 年 5 月推出了 x402,它使用该状态码,让服务器可以报价,Agent 则能在同一次 Web 请求中完成支付,无需账户或注册。Cloudflare 于 2025 年 7 月推出 Pay Per Crawl,让网站能够向 AI 爬虫收费。此后,它又转向在内容出现在 AI 回答中时向出版商付款。

这些机制大多是在讨论 AI 公司如何大规模为内容付费。而我想要的略有不同,并且我认为它更具人情味:由我自己的助手花费我自己的预算,向那些作品帮助过我的具体的人付款。

为什么这让我兴奋

这类事情让我对 AI 感到真正兴奋,因为为他人提供的价值付费,是一件非常人性的事,而管理实现这一过程的基础设施始终是难题。

在我看来,多个事情正在同时发生,但它们有一个共同主题:低摩擦、细粒度的价值交换——以数字方式完成。对我来说,这是一种人类发展,而不是技术发展。

Thinking About Different Types of Digital Value Exchange (2021)

打赏罐、Patreon、Flattr 和付费新闻通讯,都要求人类自己处理记账工作;而助手几乎就是为记账而生的。

我非常乐观地看待这一选项。我想我会尝试围绕它做些东西,可能先从我自己的网站和我自己的助手开始;但如果你知道任何正在向这一方向发展的系统,请告诉我。

注释

English / 原文

A charcoal sketch of a giant purple robotic hand pouring streams of small sienna coins down to four human creators standing on a rock ledge beside the things they made: a sketch on an easel, a stack of handwritten pages, an open book, and a lute

You know what I find super exciting about all this agentic stuff? Authors being able to publish a request for payment right on their content, along with an interface for submitting micropayments.

I've wanted this for a long time. I wrote about it in 2015 in Social Micropayments, and then again in 2016, when I described a Digital Assistant that watches what you enjoy and pays the creators out of a monthly budget. I'd say my thinking on it is about the same as it was back then. What's new is that the agents—the real ones, the kind I run all day in my own setup—are actually here, and they're getting good.

Flattr had the right idea

There was a system a long time ago called Flattr. It was started in 2010 by Peter Sunde (one of The Pirate Bay founders) and Linus Olsson, and it worked by people setting up a Flattr account beforehand, which is problem number one. If the site you navigated to also had Flattr, you could hit a Flattr button, and it would take a percentage of whatever you had in your monthly Flattr budget and send it to that person.

So if you had a $10 budget and you Flattr'd ten articles that month, each author got a dollar. If you only liked one thing, that one author got the whole $10. I loved that model. I still think it's basically right. Flattr shut down in 2023.

The problem was friction, and it was everywhere. You needed an account, the creator needed an account, the site needed the button, and you had to remember to actually click it while you were reading. Every one of those steps is a place where people drop off, and they did.

Let the assistant do it

I think we could do the same sort of thing, but all brokered by our Digital Assistant. It would know how much we have per month to give to people, and we would have some sort of payment mechanism, which could be micropayments through Stripe or whatever platform can do it well.

Anything hitting that content has the option to pay for it with micropayments, or even larger payments if it wanted to. If a user or an agent found it extremely useful to it or to its principal, it could pay a little more.

Because agents are working 24/7 and they're able to do so many things at once, my assistant could be off doing research for me and paying out $2, $5, or $12 to the people who provided the best content for that research. And I'm out of the loop for that part. I think that's the most important part, honestly, because remembering to click was always where the whole thing fell apart.

What it would take

What's cool about this is that there aren't many manual steps involved:

  • Creators configure their sites to support a payment endpoint, basically a price or a "pay what it's worth to you" request that sits next to the content.

  • People configure their harnesses, whatever is running their agents, to support payments.

  • The user and their agent come to some sort of agreement about how much money they're willing to spend this way, per month and per item.

After that, it just happens naturally as part of agent activity. The agent reads something, uses it in whatever it's doing for me, and if that piece ended up mattering, it pays the person who wrote it. If it didn't matter, maybe it pays a few cents, or nothing, depending on what we agreed.

I'm honestly not sure yet what the right payment rail is. It could be Stripe, it could be stablecoins, it could be something that doesn't exist yet. I also don't know how creators should price things, or whether most of them would just leave it at "pay what you think it's worth" and let the agents figure it out. I kind of suspect that last one.

I'd also want to set priorities, the same way I described in 2016. Something that changes how I think about a problem should get a lot more of my budget than something that was just funny. And the budget is the guardrail. An agent that can spend money needs a hard monthly ceiling and a log I can actually read, or this turns into a security problem pretty fast.

Some of the plumbing already exists

When I posted this idea on X, I asked if anyone knew of systems working toward it, and a few pieces are already out there.

HTTP has had a 402 Payment Required status code sitting in the spec since the 1990s. It's literally marked "reserved for future use." Coinbase launched x402 in May 2025, which uses that code so a server can quote a price and an agent can pay it inside the same web request, with no account or signup. Cloudflare launched Pay Per Crawl in July 2025 so sites could charge AI crawlers. It's since moved to paying publishers when their content shows up in an AI answer.

Those are mostly about AI companies paying for content at scale. What I want is a little different, and I think it's a lot more human. It's my own assistant, spending my own budget, paying the specific people whose work helped me.

Why this gets me excited

This is the type of thing that gets me really excited about AI, because it's a very human thing to pay somebody for value that they have provided, and managing the infrastructure of doing it has always been the problem.

The way I see it there are multiple things happening at the same time, but with a common theme: low-friction, granular value exchange—done digitally. To me this is a human development, not a technological one.

Thinking About Different Types of Digital Value Exchange (2021)

Tip jars, Patreon, Flattr, and paid newsletters all asked the human to do the bookkeeping, and an assistant is pretty much built for bookkeeping.

I'm really optimistic about this as an option. I think I'm going to try to put something together around it, probably starting with my own site and my own assistant, but if you know of any systems working toward this, let me know.

Notes