In the late 1990s, Tim Berners-Lee described the World Wide Web in which “intelligent agents” perform most of the digital duties for humans. Over the decades, the Internet grew, but its operation remained manual. It was the person who clicked, filled out forms, or switched tabs. The current breakthrough is this language models are no longer just answer-providing machines and are now equipped with tools to act — that is, they become agents.
Today's web, however, is a patchwork of thousands of APIs, documentation and “dialects” of individual digital services. People don't care that much. We have learned to read instructions, test, remember exceptions to popular rules and we just somehow manage to use individual services or programs (although logging in to government websites using a bank is still a sequence of actions devoid of logic).
The next part below the video:
For an agent, the human-made Internet, which is diverse and full of workarounds and exceptions, is an integration nightmare. An AI agent “thinks” in natural language and needs a uniform way to query services like: what can you do and how can I use you safely. This is where Model Context Protocol (MCP) comes into play, an open standard that aims to standardize the connection of agents to data and tools. Instead of writing integrations every time, the service can set up an MCP server and the agent itself will ask about options, from downloading files to performing a specific action.
In parallel, the “agent-to-agent” communication layer is growing. Google proposed the Agent2Agent (A2A) protocol, which is intended to enable agents to communicate and coordinate work between different providers and platforms. So that one agent can delegate part of the task to another, agree on the data format and complete the whole thing in a safe way.
So that it does not end in a format war like in the times of the first web browsers, the industry is trying to arrange standards in a common place. In December 2025, the Agentic AI Foundation was announced under the Linux Foundation umbrella, with projects such as MCP and other pieces of infrastructure for agents. The signal is clear: The “agent” Internet is to be built on open principlesand not only in closed ecosystems of the largest players.
Check also: Many companies will not survive without AI agents. Competitors are already increasing productivity and reducing costs
The browser becomes a task worker
The most spectacular change is happening in the place we know best, the browser. This is starting to look like an operating system for agents (at least that's what it's aiming to do), and not just a website opener. OpenAI already showed Operator and later included it in ChatGPT as an agent mode – the model can use its own virtual browser and perform repetitive tasks like a human, such as clicking, scrolling and filling out forms.
The next step is ChatGPT Atlas. It's a full-fledged browser with ChatGPT inside, which is supposed to understand the context of what you see on the screen and complete tasks without copy-pasting between tabs. This is important because online tasks are in practice processes, e.g. research – selection – formalities – payment – confirmation – change in the calendar. Atlas is designed precisely for this workflow.
Perplexity, in turn, released Comet, a browser that has been around since the beginning AI-first and promotes the idea that the assistant can conduct entire browsing sessions, and the user only has to clarify the intention. This is no longer a cosmetic function to summarize the content of the page, but an attempt to turn the browser into a tool for delegating work.
However, the most groundbreaking thing is shopping, because this is where reading on the Internet ends and a financial transaction begins. In September 2025, OpenAI announced Instant Checkout in ChatGPT – the first steps towards agentic commerce, where you go from chatting to paying in the same interface. The feature is based on Agentic Commerce Protocol (ACP), a technology built in partnership with Stripe. It is to be made available as an open standard for e-shop integration.
What's getting hot is demonstrated by Amazon's conflict with Perplexity. Amazon sued the company, claiming that a Comet shopping agent was interacting with the site in a way that was inconsistent with its rules, “pretending to be human” and creating risks. However, the dispute is not about a legal detail – it is a fight over control over who will be the purchasing interface in the near future. A website and its advertisements designed with people in mind, or an agent who is supposed to cut out the entire envelope and only deliver the effect, ignoring the rest.
Read also: AI agents allow you to save and earn. This can be your strong competitive advantage
How to make money on it and what to be afraid of
If an agent has to browse and buy faster than a human, the economy of the Internet changes. So far, the web has been a monetization of human attention: SEO activities, feeds, search engine advertising. In an agent-based scenario, the importance of what can be called agent's attention increases — we no longer convince people with an advertising banner, but only with their digital representative. For example, rankings, source credibility, return conditions, price transparency and data structure improve significantly. The Economist describes this as a shift away from the internet pull (man initiates) to the Internet, where the agent increasingly often pushes ready-made actions and proposals.
This is where the “AI will help you earn” space also appears, but not as a magic button, but as new market architecture. A store that implements an agent-friendly shopping path (e.g. through ACP or its own, secure integration points) may become easier to select by other people's AI assistants. Developers and companies can build microservices and tools to agent standards – ones that an agent can discover, understand, and use safely, rather than requiring a human to dig through panels and forms.
Open infrastructure such as MCP and initiatives under the Linux Foundation suggest that this ecosystem will grow like the website and plug-in market in the past. Except the client will increasingly be the algorithm itself.
There is also the other side of the coin. What? When an agent acts, he can also be manipulated. The most prominent risk is the indirect version of prompt injection, i.e. a situation when malicious instructions are hidden in the content of a page, file or HTML and intercept the behavior of an agent that has the tools to act. Research from 2025 shows that web agents may be vulnerable to such attacks if the architecture does not isolate contentand the tools have no hard restrictions on permissions.
In addition, there are classic security failures in the new standards for agents. An example is the flaw in the NLWeb project, which the media described as a serious path traversal bug – a warning signal that in the race for the future web it is easy to repeat old mistakes, only with much higher stakes, because the agent can not only read, but also perform specific actions.
Therefore, the most reasonable forecast for now is not “man disappears from the loop”, but “man changes his role”. From an operator clicking in the UI (user interface), he becomes a person who defines goals, constraints and the level of trust. Agents will get better at delivering results, but the internet for machines will force new ethics and new engineering. And later? Later we will see even greater automation.
Author: Grzegorz Kubera, journalist of Business Insider Polska
