After the "Lobster Father" complained about the human internet, finally someone took it seriously and did something about it.

（Source: Zhīqīng of Machines）

Editor｜Qiān Zhang

I don’t know if everyone still remembers—last March, AI big shot Karpathy posted a tweet. The gist was that: most of today’s content is still written for humans, but in the future, the entities reading that content might not be humans, but AI. Therefore, starting now, we need to think about how to write documents in a way that’s more friendly to AI.

Honestly, when I saw this viewpoint at the time, I hadn’t realized what it actually meant. Many netizens probably felt the same way. Some even said, “It’s still too early to think about this, after all, most people who are online right now are still human.”

But in just a year, the situation has changed. After many people experienced “Hummer” (Lobster), they were no longer willing to even do the most basic task of organizing desktop files themselves—let alone doing the dirty, tiring work of searching the internet for information.

There’s almost no doubt that Karpathy’s idea of “AI becoming the main force on the internet” will soon become reality, because humans who have tasted the sweetness can’t go back. So what’s our internet like right now? As one netizen put it—still “a gravel road in the carriage era.”

For Agents, this road is full of obstacles everywhere—every kind of verification and login program gets stuck, and tools found on the internet have to be tried one by one. Tokens are used as if they’re free (in reality, they’re very expensive). Even if the task succeeds, you still have to wait for half a day, just like dial-up internet back then.

Liu Hongtao, who has made his way through the PC internet and the mobile internet eras, told me: this kind of situation is actually hard to accept. You have to know that the human internet usability standard is 99.9%. And this group of people even once “fought” to add more “9s” to that number. But right now, the success rate of Agents calling external tools is only 60%—and that’s the result of single-step calls; with just a few more steps, it drops to below 30%.

So when “the father of Lobster,” Peter Steinberger, complained in an interview that today’s internet infrastructure is extremely “unfriendly” to Agents, Liu Hongtao immediately felt strong resonance. And the problem Peter pointed out is exactly the startup direction he had already been optimistic about the year before and officially entered in the spring of last year—Agent Internet Infra.

Peter Steinberger’s core judgment is: the current internet was not designed for agents, and it is becoming increasingly worse for them due to factors like blocking, CAPTCHAs, permission systems, and missing CLI/API, etc.; therefore, the next-generation internet/software infrastructure must be rebuilt toward being agent-friendly. (Subtitles generated by AI)

Liu Hongtao’s new company is called AgentEarth. The core team’s three people all have solid resumes. He himself served as CEO at Yunzhihui, an intelligent operations unicorn, and has experience in large-scale validation of enterprise-level infrastructure from 0 to 1. CTO Danming Hui (Lucas) was an early builder of Didi’s intelligent operations system; he has experience building and operating real-time large-scale matching systems used by hundreds of millions of people and massive ride-hailing orders. The chief scientist, Professor Xue, has also focused on national-level cutting-edge network technologies for many years, and the underlying protocol stack is his strong suit.

AgentEarth CEO

Liu Hongtao (left) and CTO Danming Hui (right)

With this kind of combination, it’s clearly not about making a simple Agent tool. As Liu Hongtao put it, what they want to do is infrastructure: at the bottom, build a high-speed logistics line for Agent Internet so that data transmission runs stably and quickly; on top, open a “boutique self-operated store”—not for people to browse, but to treat Agents as true end users. When they “walk into the store,” they can quickly call high-quality tools that have been screened and governed. The first part relies on their new-generation transmission protocols they’ve been developing for years; the second part is to make solid the whole set of tool aggregation, hosting, and intelligent orchestration—so that Agents don’t have to keep trying everywhere like headless flies, crashing into walls everywhere, saving valuable time and tokens.

As for exactly how they plan to do it, Liu Hongtao also talked in detail.

An internet designed for people

A minefield for Agents

Recently, Anthropic and OpenAI have once again popularized a term: Harness Engineering. In a blog post, Anthropic said that with the same model and the same prompts, the game that initially comes out can’t be played—but if you change the way it runs and the environment, you can get a good game.

These frontier organizations are warning everyone through experiments—while improvements to the model itself are crucial, you can’t ignore the external environment around how the model runs, otherwise it will affect the performance of large models.

This also explains why OpenAI claimed as early as 2024 that large models had reached the level of doctoral ability in some areas, but it wasn’t until this year that the production side started to feel it in a tangible way.

Building this environment is far more complex than you might imagine. In the past one or two years, engineers in the Agent Infra field have been trying to solve issues like long-term memory storage and execution orchestration, providing a foundational support system for Agents to run stably. But this wave of “Lobster fever” has also fully exposed a shortcoming—external calls. You have to know that even for a simple ticket-booking operation, an Agent has to call a dozen or more external tools. So when Agents start “handling business online” like humans, the network layer becomes the new battlefield.

Liu Hongtao mentioned that when facing this new battlefield, the infrastructure build-out must follow a new logic, because an Agent’s online behavior is radically different from humans.

Human browsing is opening a browser, searching keywords, then clicking the webpages they’re interested in. After that comes browsing, thinking, and judging. People tend to stay on a single webpage for a longer time, but the overall browsing behavior isn’t that complex. Plus, cache technologies like CDNs (caching once to serve a large group of people) can ensure speed. Well-crafted UI designs can improve efficiency, and various tools have been used for years and are relatively smooth.

But an Agent is different. It doesn’t go online to “look.” It goes online to “get the job done.” The tools it needs for a single task span multiple models and platforms, and the execution chain is long. If anything gets stuck in one place, the whole task falls into a trial-and-error black hole. Also, it actually has higher requirements for speed than humans do, because it doesn’t need reaction time—it just wants results as fast as possible so it can move immediately to the next step.

However, the reality is that most webpages and tools on the internet are still designed for people (like the “you’re not a robot” verification for Agents mentioned by Peter in the interview). They haven’t been curated and adapted specifically for Agents, so the long chains in Agents are easy to break. Moreover, some of the things an Agent gets from browsing are only what it itself needs (for example, generating a certain image). Once it’s used, other people can’t use it, so CDNs fail and speed can’t really pick up.

When these characteristics stack up, the human internet infrastructure starts to “not fit” in front of Agents. And right now, Agent Internet is still in a savage, early growth stage. External tools are a mixed bag, interfaces are chaotic, and quality varies widely. During the calling process, Agents frequently “lose their sense.” They burn a huge number of tokens in repeated trial-and-error and repeated context passing, and task completion speed can’t be improved either.

At this point, what needs to be done for Agent Internet Infra is very clear: it aims to provide a foundational network protocol and middleware system that allows a massive number of agents to autonomously discover, securely connect, and collaborate in a trustworthy way. It’s dedicated to solving how Agents connect with external systems, and how Agents collaborate seamlessly with each other just like humans use the internet. Its core capabilities include identity authentication, communication protocols, permission governance, cross-platform tool calling, data transmission optimization, transaction payments, security management, and more.

Currently, some companies have started pushing in this direction. For instance, Cloudflare released a Markdown for Agents that makes it easy for Agents to read webpages, and Google released WebMCP that connects browser environments with local compute resources… But overall, this direction is still at an early stage, and new-generation Agent Internet Infra service providers are still missing.

An internet for Agents

How to save money and save time?

In the Agent Internet Infra direction, the startup logic of Liu Hongtao and others has one core anchor: from day one, treat Agents as the primary subject users of the network—that is, end users (in the past, it was assumed to be humans). This aligns with Karpathy’s judgment.

Once you anchor this assumption, the direction of network infrastructure optimization shifts from “serving human experience” to “serving task completion rate and completion efficiency,” from “platform provides connections” to “platform is responsible for outcomes.” In other words, they mainly consider: can your “Lobster” complete tasks with high-quality, high reliability, and high efficiency by leveraging my platform? I need to be responsible for your results. I need to save you money and time.

Most importantly, this isn’t something that stays at the conceptual level—it has been turned into product decisions.

The most obvious point is that they deliberately don’t build interfaces for people, don’t do complicated developer experiences, and instead only build standardized Agent interfaces. Behind this is actually a very firm judgment: in the future, it won’t be developers configuring tools—it will be Agents assembling tools themselves. If you believe this, then all the layers designed for “ease of operation for humans” are just short-term transitions.

So how do they turn “high quality and high reliability” into a differentiator? Here, it actually breaks down into three layers of the tech stack.

In the middle layer, they move “tool quality problems” from the Agent side to the platform side. The mainstream approach right now is to let the Agent pick tools and do trial-and-error, filling the gaps with more tokens. The result is high cost, low success rate, and uncontrollability. In this layer, they take over that job and provide Agents with a single “gateway” to access external services. That means Agents don’t need to know which tools are good— the platform has already selected and covered them. If something fails, it switches immediately. Billing is also unified here. Everything is transparent with the data, so people behind the Agent can see which tools were used and how many times calls were made, and tokens are clearly accounted for—no more black hole that swallows money.

The layer above that is dedicated to ensuring early quality with a “self-operated” logic. At the beginning, they didn’t open an ecosystem; they selected tools themselves, emphasizing stability, efficiency, and high quality—like the early JD self-operated mall, where the core was to help “Lobster” users complete high-quality tasks. After generating traffic, they will also open third-party merchants to join, and they adopt a highly intelligent process using a tool recommendation algorithm based on large models and call-optimization strategies.

The bottom layer is also their most hardcore: they sink “reliability” down to the transmission layer, using their own integrated transmission-storage-computation scheduling protocol to speed up underlying data transfer.

In real-world testing, this protocol is 2–10 times faster than the best open-source protocol currently in the industry—Google QUIC. In recent tests, it has even reached more than ten times. That means if your Agent wants to transfer files, images, or video from a remote location—especially those kinds of personalized content generated just moments before—this protocol is much faster than traditional methods.

People in the industry may already know this: protocol is a whole self-consistent set of rule systems, so protocol R&D isn’t a short-term job—you can’t split it into modules and push in parallel like writing an APP. Building a new protocol is like raising a new species: you start with a seed and grow it slowly in a specific order. Each step has to wait until the previous part is fully set before you can move on. Even if you add more engineers, it doesn’t compress the time it takes to “grow up.” And the implicit knowledge inside protocol design—edge cases of network behavior, and the pitfalls they’ve stepped on—can only come from long-term accumulation. Liu Hongtao said their protocol wasn’t built overnight either; the development cycle is measured in ten years. The earliest experience accumulation was originally to optimize TCP/IP, but it turned into the company’s core technical moat now.

The ceiling of this

Might be higher than you think

In the PC internet and mobile internet eras, the growth in the number of netizens and the time each netizen spends online are often seen as the core drivers of total market growth. But as both approach their limits, this kind of growth has already hit a ceiling.

The emergence of this new track—Agent Internet—is rewriting the rules of the game. A company, or even a single person, can deploy hundreds or thousands of Agents. One Agent can run multiple tasks at the same time, and these Agents don’t need to sleep. This means that the upper limit of the traffic and value carried by Agent Internet Infra is currently hard to estimate.

This also means that this layer is likely to spawn a new batch of big companies. Looking back at the PC internet and mobile internet, almost every layer of infrastructure eventually produced independent companies, because the problems were general enough and the demand was hard enough—so sooner or later someone would turn it into a platform. Agent Internet is no different, and this time the user scale and call intensity are even more extreme. Many foundational problems are still blank, leaving more room.

In this early stage of the race, AgentEarth has already occupied a fairly good position.

On one hand, they made the judgment early and decisively. From the start, they built the system around the idea that Agents are users, focusing on Agents completing tasks with high reliability and high quality. On the other hand, the team structure is relatively uncommon. Low-level protocol capability is very hard to catch up with in a short time, and the people who have fought in scenarios like “hundreds of millions of users and large-scale real-time matching of massive resources” are also rare. Systems with requirements this extreme for stability, efficiency, and fault tolerance usually don’t get many chances in daily work to “train” like this. Once the calling scale of Agents grows, this experience becomes extremely valuable—and it can’t be made up for in a short time.

Just yesterday, AgentEarth released a beta version of their product and began small-scale testing. The test link is as follows: Agentearth.ai

Interested readers can also scan the QR code to join the group and exchange experiences:

Abundant information, precise analysis—available in the Sina Finance app