AI's 'Steamroller Effect' and the 'Workshopization' of Software Herald the Agentic Internet

Recently, SpaceX and Meta have successively launched AI agents, OpenAI and Anthropic have cut model prices, blockchain payments have become a key element of the agentic internet, and AI agents are reshaping the software economy and the form of Web 3.0.

Author: Meng Yan's Blockchain Thoughts

Some very interesting things have happened over the past month or so.

On Aug. 11, SpaceX released an AI agent application called Grok Bot. It is an "AI colleague" that works around the clock on its own cloud computer.

On Sept. 8, Meta launched a personal agent application called Muse. It can open browsers, fill out forms, and bargain on behalf of users. Within ten days of launch, it topped the US App Store charts and triggered a sell-off of traditional online intermediary platform companies, such as Expedia, Airbnb, and Booking, whose stocks plunged for several consecutive days.

Also on Sept. 22, Anthropic, the maker of Claude, released a new model, Opus 5.5. Not long before, OpenAI, the maker of ChatGPT, launched two new versions of GPT-6, both priced at half of the previous generation.

In mid-September, Salesforce, one of the world's largest SaaS software companies, held its annual conference. Patrick Stokes, president of its applications business, said that AI will take apart software interfaces and then replace them.

These events come from different companies and industries, concentrated within two months. They seem like separate moves, but underneath they are the same thing: AI agents that can get things done for people are starting to interact with software and the internet on people's behalf. The shape of the next-generation internet is emerging.

1. The Steamroller Effect

The head of an AI incubator told me that over the past year, the biggest wake-up call for AI entrepreneurs has been discovering that AI, unlike the internet and blockchain, is an arena of extreme centralization and extremely rapid centralization. Startups have a very short survival window and must move fast, seek acquisition, and find a buyer without hesitation before being crushed by the giants, cashing out and leaving. As for the dream of growing big on their own, forget it.

This is the mindset of the grass in front of a steamroller.

The "steamroller effect" is a source of anxiety for many AI entrepreneurs. Every step forward by frontier models unknowingly crushes a batch of startups. A few listed companies can at least lament through plunging stock prices, but the instant disappearance of hundreds or thousands of small teams goes unnoticed.

First, look at the money. According to PitchBook, a US venture capital data provider, global AI venture investment hit a record in the first half of 2026, with more than half flowing to OpenAI and Anthropic. It is like a banquet with thousands of tables, where half the dishes go to the head table, and only two people sit at that table.

Now look at specific companies. Google has an AI note-taking product called NotebookLM. Users put documents and web pages into it, and it turns the content into podcast-style audio explanations. It went viral online in 2024. Its head, Raiza Martin, left Google with two colleagues to build an AI podcast app called Huxe, which generates daily audio briefings based on users' emails and calendars. Google Chief Scientist Jeff Dean invested in it.

On May 21 this year, Spotify, the world's largest music streaming platform, released an update that included similar features. On May 22, Huxe announced it was shutting down.

The same applies to listed companies. In February, Anthropic released a set of industry plugins for Claude covering legal, financial, and other sectors. Thomson Reuters, the parent company of Reuters, which makes its living selling legal, tax, and accounting information, saw its stock drop more than 15% that day.

After Claude released the Opus 5.5 model, a large number of users posted on Twitter the exquisite videos they made with simple prompts, cheering with delight, while teams that had toiled countless sleepless nights in the field of AI video production could only slump before their screens and weep silently.

Heaven and earth are not benevolent; they treat all things as straw dogs. OpenAI, Anthropic, and other giants are not out to get you. They bear you no ill will, have no intention of competing with you, and may not even know you exist. They simply move forward, and then you are reduced to ashes.

How does that tired old meme go? Destroying you has nothing to do with you.

2. The Workshopization of the Software Economy

In February this year, Claude Opus 4.6 was released. With the support of this model, Claude Code suddenly became so powerful that people who cannot write code at all suddenly felt their tech dreams could be revived.

More and more people are making software. There is an AI platform called Lovable that lets people who cannot program create websites and apps through conversation. It claims that the platform adds 1 million new projects every week, with users mainly being programming novices who cannot write a single line of code.

On the other hand, selling software has become tough. The main index fund tracking US software stocks fell more than 24% in the first quarter of this year, the worst quarter since the 2008 financial crisis. Stock prices reflect expectations and also factor in high interest rates, so this does not mean demand has already shrunk. In September, the hammer fell on internet intermediaries, as the market repriced software sold to humans.

I call this the "workshopization" of the software economy. The Industrial Revolution moved spinning and weaving from homes into factories; AI is moving software development from factories back into homes. Every individual and every small company can set up their own software workshop.

In the past, opening a small shop and needing an inventory system meant finding an outsourcing company, discussing requirements, signing contracts, and waiting for delivery. Now, you just talk to a computer, and if things go well, you can have a prototype in an afternoon. The same goes for internal tools like customer management and staff scheduling.

Can the next big company grow out of these workshops? It is very difficult. The problem with workshops is isolation, and the more workshops there are, the more obvious it becomes. A small country with few people: you can hear the cocks crowing and dogs barking next door, but people grow old and die without visiting each other. Laozi regarded this as an ideal, but in software it is a predicament: I cannot use your product, and you should not even think about touching my data. What is made in a workshop can only be used by the maker.

Never have there been so many people making software, and never has it been so hard to sell software.

3. The Internet of Agents

AI is extremely centralized, while software is moving toward decentralization. Fire and ice, but they are two sides of the same coin. It is precisely because AI capabilities keep strengthening that the space for software is being squeezed.

This is only the beginning. People still use AI to develop software and then use the software. A few steps further, and software will not even be needed. Users will interact directly with AI agents to solve all problems.

Muse is currently the clearest prototype.

Meta assigns each user a virtual computer in the cloud, and Muse works on that computer. When the user closes the app, it keeps working, notifying the user when a task is done or needs approval. If the other party has an interface, it calls it directly through connectors; if not, it opens a browser and operates page by page like a human.

Filling out forms, comparing prices, negotiating prices—tasks that users used to do themselves are now handed over to it.

It is still not very good at this. PYMNTS, a US payments media outlet, asked it to do three things: restock toilet paper on Amazon, order a Domino's pizza, and book a restaurant reservation. It failed at all three. The reviewer said that for something a human can do in 30 seconds, it added an extra layer of management.

But what we need to look at is the direction. How is this different from using apps in the past? In the past, booking a flight meant opening several travel websites yourself, comparing prices one by one, and filling in your name and ID number field by field. In the future, you just tell Muse, "I'm going to Tokyo next month, find the cheapest direct flight," and it handles the rest.

For the first time, users speak to a single agent, and it deals with the entire internet on their behalf. Apps and websites recede into the background.

This is not just a change in how people interact with the internet; it is another upgrade of the internet itself.

Looking at the history of the internet: Web 1.0 connected documents—it was an internet of documents, which people read; Web 2.0 connected applications and services—it was an internet of applications, which people used. What does Web 3.0 connect?

The term Web 3.0 has had several interpretations over the past twenty-plus years. Berners-Lee, the inventor of the World Wide Web, systematically articulated the "Semantic Web" in 2001, aiming to tag web pages with labels that machines can read. The blockchain community proposed the "internet of value," aiming to make money and assets flow online like information. Both descriptions captured features but neither described the form.

Now it is clear: the form of Web 3.0 is the internet of agents. Agents read, use, and negotiate on behalf of people. One agent deals with hundreds of apps and thousands of APIs for a person. The human experience no longer matters; what matters is that the agent finds it good, and then it is good for both. Over the past twenty years, product managers and designers built products around the human experience, competing on interfaces and operations. Going forward, products must be written for agents to read, competing on clear interfaces, clean data, and readable terms.

Salesforce launched a set of interface-free toolkits, opening its entire platform to agents through MCP and APIs, and even built its own features directly into Claude. A company that sells software through interfaces has dismantled its own interface.

The CEO of Expedia, an online travel platform, summarized the company's new strategy as needing to "be everywhere agents are." Travel platforms used to pull people to their websites; now they must go to the agents.

Who is this bad news for? Wall Street is repricing "consumer inertia." Many businesses rely on customers being too lazy to compare prices, switch apps, or call to haggle. Agents do not mind the hassle; they will do all of these things. That is why the market is selling off such businesses.

Even Meta itself must transform. The company built its fortune selling users' attention, with revenue mainly from advertising. Agents do not watch ads or scroll feeds. The revenue source Zuckerberg has found for Muse is taking a small cut from transactions.

Thirty years ago, websites began optimizing for search engines, which later became a business. This time, the entity to adapt to has changed from a ranking machine to an agent that makes decisions for its owner. Products that agents cannot understand are effectively nonexistent; products that agents can understand and use smoothly will win business. The same applies to capabilities made in workshops: if agents cannot understand them, they cannot be sold.

Ben Thompson, a US tech analyst, said that agents will become the "ultimate gatekeepers." Whoever controls the agents controls user demand.

Another thing that will definitely happen is blockchain payments and token economics. Agents exchange value with each other in small amounts, many times, with unfamiliar counterparties. Whether to pay and how much is decided by programs on the spot, with no human oversight. Card payments charge 2.9% plus 30 cents per transaction; paying 1 cent would incur a fee 30 times the amount. Card networks cannot handle such transactions.

Visa's own research also admits that cards cannot handle this segment, assigning machine-to-machine micropayments to stablecoins. Stablecoins are digital currencies typically pegged one-to-one to the US dollar and circulating on blockchains, with a total supply exceeding $300 billion.

My judgment is that blockchain-based digital payments and token economics—that is, a set of arrangements using on-chain credentials for pricing, settlement, and revenue distribution—are an essential foundational element of the agentic internet.

I have always remembered a meeting at the Digital Asset Research Institute in 2018, where Mr. Zhu Jiaming said that in the long run, blockchain is not for people; it is for AI. I thought it was very insightful at the time, but I could not figure out how AI would use it. Now it is clear.

4. "AI Agent-Friendly" Jobs

The impact of AI on jobs has always been a hot-button topic. In the AI era, what kind of people can survive and thrive in the long term? Or, to put it bluntly, what jobs can avoid being crushed by the steamroller?

They must be jobs that collaborate well with AI agents.

Top AI scientists at large model companies certainly qualify, but there are very few of them, the barrier to entry is extremely high, and it is not accessible to most people.

The recently popular role of Forward Deployed Engineer (FDE) also seems to have fizzled out quickly. People soon realized it was just a trendy name for on-site outsourcing, and crucially, it was often a one-time gig. All kinds of imagined AI enterprise solutions, when rushed into implementation, turned out to have such poor data quality that the deployed enterprise AI was like the village idiot. You had to start from scratch with the underlying data, but isn't that what data engineering is supposed to do? As a result, FDE became "Fooled Data Engineer." Who wants to do that lousy job?

Currently hot are AI engineers building products on AI models and AI-Assistant Developers using AI to assist in developing traditional software. But if the two judgments—that AI agents will squeeze software and that the internet is moving toward agents—are correct, then these two job roles will also transform. In the past, products were made for people; increasingly, they will be made for agents. In the past, optimization was for search engines; in the future, it will be for agents.

There is another type of work that is more universal: managing AI agent organizations.

When agents become powerful enough, you won't even need to develop software. As long as you can manage them well, set requirements, and make judgments and decisions at key points, you can accomplish the vast majority of work. But that is far from enough. In the future, you will need to design and organize several, dozens, hundreds, or even thousands of agents, forming them into an efficient and formidable legion, designing efficient processes, ensuring security, controlling expenses and token budgets, effectively evaluating performance and continuously improving, and then competing and fighting against your competitors' agent legions of equal scale.

For example, a company's procurement agent must simultaneously negotiate prices with the sales agents of a thousand suppliers, and the opposing agents are equally smart. Victory depends on how this side divides labor, delegates authority, and assigns accountability. Management must become engineering: permissions, boundaries, incentives, and audits must all be trainable and measurable. This is a new management science and a new systems engineering.

Conway's Law states that organizations that design systems will ultimately produce systems that replicate their own communication structures. In the agent era, the members of the organization are agents, and the system and the organization become one. The organization is what the agent system is; designing the agent system is designing the organization.

This is not the idle fantasy of a few individuals. Microsoft said in its annual Work Trend Index report that everyone will become a boss of agents; NVIDIA CEO Jensen Huang said that IT departments will become the HR departments for AI agents.

Therefore, the most valuable jobs in the agentic internet are, first, AI engineers who can build products and services for agents, and second, legion commanders who can lead a thousand agents to defeat another thousand agents.