A cluster of announcements from major tech companies over the past two months points to the same underlying shift: AI agents are starting to handle software and internet interactions on behalf of users.
On Aug. 11, SpaceX released Grok Bot, described as an AI coworker that works around the clock on its own cloud computer. On Sept. 8, Meta launched Muse, a personal agent that can open a browser, fill out forms and bargain for users. The article says Muse climbed to the top of the U.S. App Store within 10 days and touched off a selloff in traditional online intermediary platforms, with Expedia, Airbnb and Booking all falling over multiple sessions.
Then on Sept. 22, Anthropic released Opus 5.5. Shortly before that, OpenAI introduced two new versions of GPT-6 at half the price of the previous generation. In mid-September, Salesforce held its annual conference, where Patrick Stokes, president in charge of applications, said AI will break apart software interfaces and replace them.
Taken separately, those events came from different companies and different corners of the industry. Taken together, the article argues, they describe one thing: agents that can get work done are beginning to deal with software and the internet for people. That is why the shape of a new internet is starting to come into view.
The “steamroller effect” hitting AI startups
The article quotes the head of an AI incubator as saying that over the past year many founders have woken up to the same reality: AI is not like the internet or blockchain. It is an extremely centralized arena, and it centralizes at high speed. Startups have a short survival window, which means they need to move fast and sell fast, ideally before a large platform crushes them. Building a large independent company, in that framing, is no longer the default dream.
The author calls this the “steamroller effect.” Every time a frontier model steps forward, it can wipe out a batch of startups without even knowing they exist. Public companies at least register the damage through steep stock-price drops. For hundreds or thousands of smaller teams, the collapse often passes unnoticed.
The money flow shows the concentration first. Citing PitchBook, the article says global AI venture investment hit a record in the first half of 2026, with more than half of that capital going to OpenAI and Anthropic alone. The author compares it to a banquet with thousands of tables, where more than half the dishes are placed on the head table and only two people are seated there.
The company examples are even sharper. Google’s NotebookLM became widely known in 2024 for turning files and web pages into podcast-style audio explainers. Its lead, Raiza Martin, later left Google with two colleagues to build Huxe, an AI podcast app that generated daily audio briefings from users’ emails and calendars. Google chief scientist Jeff Dean invested in the company.
On May 21 this year, Spotify rolled out an update that included a similar feature. On May 22, Huxe said it was shutting down.
The same pattern showed up in listed companies. In February, Anthropic rolled out a set of Claude plugins for legal, financial and other professional use cases. Thomson Reuters, whose parent company is Reuters and whose business depends on selling legal, tax and accounting information, fell more than 15% that day.
The article also points to the reaction after Claude released Opus 5.5. A large number of users posted polished videos on Twitter that they had made using simple prompts. For teams that had spent long stretches building products in AI video creation, stronger model capability suddenly changed the ground beneath them.
The author’s point is not that companies such as OpenAI and Anthropic are deliberately trying to kill every startup. Quite the opposite. They may have no hostility, no direct intent to compete and no idea a small team even exists. They move forward on their own path, and a layer of businesses disappears as a result.
Software becomes easier to make, but harder to sell
The article then turns to software and argues that AI is pushing the software economy toward what it calls a workshop model.
In February, Claude Opus 4.6 was released. The article says Claude Code became so capable on top of that model that even people who could not code began to feel their technical ambitions might still be salvageable.
That lowers the barrier to building software. The piece cites Lovable, an AI platform that lets non-programmers create websites and apps through chat. By the platform’s own account, it is adding 1 million new projects a week, and most of its users are people with little to no coding experience.
At the same time, life has become tougher for software sellers. The article says a major index fund tracking U.S. software stocks fell more than 24% in the first quarter, the worst quarter since the 2008 financial crisis. The author adds that stock prices reflect expectations and also include the effect of high interest rates, so they do not prove that demand has already contracted. Still, by September, the market had started hitting online intermediaries, which the article reads as a repricing of software and platform businesses built for human users.
The workshop analogy is central to the argument. The industrial revolution moved spinning and weaving from households into factories. AI, the author says, is taking software creation from factories back into households. Every person and every small company can now run a small software workshop of its own.
The article gives a simple example. A small shop that once needed to hire an outsourcing firm, negotiate requirements, sign contracts and wait through a delivery cycle to get an inventory system can now describe the need in plain language and produce a rough version in one afternoon. Internal tools for customer management or staff scheduling are changing in the same way.
But the author does not present that as a path to the next giant software company. The problem with workshops is isolation, and that problem grows as workshops multiply. One team’s product is often usable only inside that team’s own environment, while its data stays off-limits to everyone else. What gets built is increasingly custom and inward-facing rather than broadly reusable.
The conclusion is blunt: there have never been more people making software, and there have never been more difficulties in selling software.
An internet where agents act for users
The article frames the centralization of AI and the fragmentation of software as two sides of the same development. As AI capabilities get stronger, they keep squeezing the space occupied by traditional software.
For now, people still use AI to build software and then use that software. A few steps later, the article argues, software itself may retreat into the background. Users will interact directly with AI agents, and the agents will do the actual work.
Muse is presented as the clearest early model. Meta assigns each user a virtual computer in the cloud, and Muse works on that machine. When the user closes the app, the system keeps going. It notifies the user when the job is done or when approval is needed. If a service offers an interface, Muse can call it through a connector. If not, it opens a browser and clicks through pages the way a human would.
Forms, price comparisons and price negotiations, tasks that once sat squarely with the user, are handed to the agent instead.
The article does not say the product is already reliable. It cites a PYMNTS test in which Muse was asked to do three things: restock toilet paper on Amazon, order a Domino’s pizza and book a restaurant table. It failed at all three. The reviewer said that for jobs a person could finish in 30 seconds, the system added another management layer instead of removing friction.
Still, the author says the direction matters more than the current scorecard. Booking a flight used to mean opening several travel sites, comparing fares and filling in names and ID details by hand. In the proposed future, the user tells Muse to find the cheapest direct flight to Tokyo next month, and the rest is handled in the background.
That means the user speaks to one agent first, and the agent deals with the broader internet after that. Apps and websites do not disappear, but they move behind the curtain.
The article places that shift in the longer history of the web. Web 1.0 connected documents, and people read them. Web 2.0 connected applications and services, and people used them. The article’s answer to what comes next is direct: Web 3.0 takes the form of an internet of agents.
It notes that “Web 3.0” has been used in several ways over the past two decades. In 2001, Berners-Lee laid out the semantic web, a plan to attach machine-readable labels to web pages. The blockchain industry later popularized the idea of an internet of value, where money and assets could move online the way information does. In the author’s view, both descriptions captured features, but neither fully described the form.
Now the form looks clearer. Agents will read, use and negotiate on behalf of people. One agent will interact with hundreds of apps and thousands of APIs. Human-facing experience will no longer be the only thing that matters. What matters just as much is whether the agent can understand the product, work through the interface and complete the task cleanly.
That leads to a different product logic. For the past 20 years, product managers and designers have focused on interfaces and user flows built around human experience. In the period ahead, products will also need to be legible to agents. Clean APIs, tidy data and machine-readable terms become more important.
The article points to Salesforce as one example. The company introduced a toolkit with no interface, opened its platform to agents through MCP and APIs, and plugged its own functions directly into Claude. A company that sold software through interfaces is now breaking those interfaces apart itself.
Expedia offers another example. The article says the company’s CEO summarized its new strategy as a need to “show up wherever the agents are.” Travel platforms once tried to pull users onto their own sites. Now they need to insert themselves into agent workflows.
That is bad news, the author argues, for businesses built on consumer inertia. Many models depend on users not wanting to compare prices, switch apps or make a phone call to negotiate. Agents do not mind any of that. They will make the comparisons and place the calls, so markets are repricing those businesses.
Even Meta has to adjust. The company built its business by selling user attention through advertising. Agents do not watch ads and do not scroll feeds. According to the article, Mark Zuckerberg’s revenue idea for Muse is to take a small fee from transactions.
The author compares the shift to search-engine optimization. Thirty years ago, websites began adapting themselves for search engines, and an entire industry eventually grew around that. This time, the thing products must adapt to is not a ranking machine but an agent that makes decisions for its user. If an agent cannot understand a product, that product might as well not exist. If an agent can understand it and use it smoothly, business follows. The same logic applies to capabilities built in small software workshops. If agents cannot read them, they cannot be sold.
The article quotes U.S. tech analyst Ben Thompson as saying that agents will become the “ultimate gatekeepers.” Whoever controls the agent controls user demand.
Why blockchain payments and tokens matter here
The piece argues that another development follows almost inevitably from an agent-driven internet: blockchain payments and token economics.
In transactions between agents, value exchange is likely to involve small payments, high frequency, unfamiliar counterparties and execution rules set by software in real time. The article says card networks are not built for that structure.
Its example is simple. Card payments often charge 2.9% plus 30 cents. If the payment itself is 1 cent, the fee makes the transaction impossible. In that sense, card rails cannot support machine-to-machine micropayments.
The author says Visa’s own research acknowledges that cards cannot handle this area and effectively leaves machine micropayments to stablecoins. The article explains stablecoins as digital currencies, usually pegged 1:1 to the U.S. dollar and circulating on blockchains, and says their aggregate supply has passed $300 billion.
From there, the author offers a clear judgment: blockchain-based digital payments and token economics — meaning arrangements that use onchain credentials to price, settle and distribute revenue — will be indispensable basic elements of the internet of agents.
The article closes this section with a personal recollection. At a 2018 meeting hosted by the Digital Assets Research Institute, Zhu Jiaming said that in the long run blockchain would not be for people, but for AI. The author says that idea felt right at the time, even if the mechanism was not yet obvious. In the current framework, it has become easier to see.
What kinds of jobs fit an agent era
The final section turns to work. The article says the pressure AI places on jobs remains an unsettled and heated issue, but it takes one position clearly: the roles with long-term durability will be those that can work well with AI agents.
Top AI scientists inside frontier model companies obviously qualify, but there are very few of them and the bar is very high.
The article also says the frontend deployment engineer, or FDE, briefly became a hot role and then cooled quickly. In practice, it often turned out to be a fashionable label for on-site outsourcing, and often for one-off projects. Companies rushing to build internal AI systems discovered that poor data quality made those deployments ineffective, forcing teams back into foundational data cleanup. At that point, the role looked a lot more like data engineering than anything new. The article jokes about the title as “Fooled Data Engineer.”
What is in demand now, it says, are AI engineers building products on top of models and AI-assistant developers using AI to build traditional software. Even those jobs, however, will evolve if two judgments hold: that agents will squeeze software, and that the internet will shift toward agents. Product work that was once designed for people will increasingly be designed for agents. Search optimization gives way to agent optimization.
The author then points to a broader category: the management of organizations made up of AI agents.
Once agents become strong enough, the article argues, companies may not need to develop as much software at all. If a team can manage agents, set requirements and make decisions at critical points, that may be enough to complete most work. But that is only the first step. The harder task is to design and organize several, dozens, hundreds or even thousands of agents into an effective system, while keeping it secure, controlling spending and token budgets, and measuring performance for continuous improvement.
The article uses a purchasing example. A company’s procurement agent may negotiate prices with 1,000 supplier sales agents, each of them equally capable. In that situation, the outcome depends not on the strength of a single agent but on the internal design of the system: how work is split up, what authority is delegated and how accountability is assigned. Management becomes an engineering problem. Permissions, boundaries, incentives and audits all need to be trainable and measurable.
The author describes that as a new kind of management science and a new systems engineering discipline. The article also invokes Conway’s Law: organizations design systems that mirror their own communication structures. In an agent era, the members of the organization may themselves be agents, which means system and organization begin to merge. The shape of the organization becomes the shape of the agent system. Designing an agent system becomes a form of organizational design.
The article says this is not just one person’s thought experiment. Microsoft wrote in its annual workplace trends report that everyone will become the boss of agents. Nvidia CEO Jensen Huang has said the IT department will become the HR department for AI agents.
On that basis, the author reduces the most valuable work in an internet of agents to two categories: AI engineers who can build products and services for agents, and organizers who can command a thousand agents against another thousand-agent system.