The internet we use today was built for eyes, clicks, and scrolling, but the next era of the web is being built for something entirely different: agents.
The Simple Version
Imagine you are trying to order a pizza. For the last twenty years, you've had to open a browser, find a website, navigate a menu, click buttons, and deal with pop-ups.
This is the "browser tab" model—a visual interface designed specifically for human eyes and hands. But what if you didn't need to see the menu at all?
What if you could just tell a digital assistant, "Get me a large pepperoni," and it handled the entire transaction behind the scenes?
How It Actually Works
The Failure of llms.txt
Currently, developers are trying to help AI agents by creating llms.txt files. These are essentially instruction manuals sitting on a website, telling an agent, "Here is how you should talk to me."
However, Liad Yosef's research shows a massive disconnect. Even though roughly 50% of the websites they tested had these files, almost none of the agents actually read them.
Instead, agents default to old-school behaviors: they crawl the documentation or the homepage, essentially trying to "read" the site like a human would. This proves that we can't just write specifications for agents and hope they follow them.
The Rise of MCP Apps
This is where the Model Context Protocol (MCP) Apps spec comes in. Rather than flattening a complex website into a wall of text that an LLM has to struggle to parse, MCP allows a server to hand a real, functional interface back to the chat.
Instead of an agent reading a description of a button, the agent receives the actual capability of that button as a tool it can invoke. It turns a static webpage into a set of actionable components.
Technical Deep Dive
The fundamental shift Yosef describes is the move toward a "nearly headless web." In traditional web development, the "head" is the User Interface (UI) that renders HTML and CSS for a human. In an agentic world, the UI becomes an afterthought or is bypassed entirely.
Yosef suggests breaking interfaces into "atoms." In a standard web architecture, a dashboard is a monolithic collection of visual elements. In an agentic architecture, these elements are decoupled into discrete, functional units that an assistant can compose.
This has massive implications for brand loyalty and discovery. Yosef shares a striking example: his team's coding agent recommended a new analytics service over their decade-long vendor.
The agent didn't care about the vendor's polished UI, their marketing, or their brand history. It chose the new service simply because it had a superior server and API integration via MCP. For an agent, the "best" product is the one that is easiest to programmatically interact with, not the one with the prettiest dashboard.
Why This Matters
We are witnessing the end of the "browser as the primary interface." When agents become the entry point for most digital tasks, the websites that force a user to open a browser tab will be punished by the very tools meant to assist them.
This shift turns accessibility and agent-readiness into the same job. If a service is easy for a screen reader to navigate, it is likely easier for an agent to navigate. If it is built with robust, atomic APIs, it becomes a first-class citizen in the agentic economy.
The companies that win the next decade won't be the ones with the best marketing or the most intuitive UI for humans; they will be the ones with the most seamless, discoverable, and functional interfaces for the machines that do our work for us.