AI Agents Will Deeply Transform Our Experience With the Web

The rise of AI Agents is crazy. Every software application, even documentation, is shipping an AI Agent or AI-powered feature.

Notion, with its 3.0 release, Cursor documentation, IDEs like VS Code with Copilot Agent and all its forks, even n8n, with workflows automation with AI Agents, are shipping Agents.

Agents, Agents, Agents. Everywhere, every time, for everything. From my point of view because the filter bubble from social media heavily distorts the perception of reality.

We no longer browse the web. We ask ChatGPT or Perplexity to do it for us, and it radically changes our relationship with software, services, and information.

We no longer do actions. We now ask for them.

This is a huge shift and the user experience is in the front line. We're not talking about changing some details on how web apps are designed; we're talking about imagining something completely new, from scratch, with no precedent.

We're talking about a complete change in usage patterns.

And this is massive. Maybe I'm totally wrong and we all laugh about it in a few years. But if this is the future, then we need to be prepared.

Function Calling Changed AI ​

In June 2023, more than two years ago, OpenAI introduced function calling capabilities in their API. This enabled AI to connect and interact with the real world. At that time, context was 16k for high-end models.

This was mainly used by coding editors to enhance their capabilities. AI could now search for information, create files, run commands, and more. This was the beginning of AI Agents. In February 2025, VS Code announced its GitHub Copilot agent mode, but software like Cursor or Windsurf had already been leveraging these capabilities for months.

MCP, which stands for Model Context Protocol, appeared in November 2024. MCP isn't directly related to AI. It's a connector between an AI application and another service to augment the AI capabilities. The AI application connects to the MCP server, lists its tools, and injects them to the AI using the function calling capabilities introduced earlier.

This is smart. It makes plugins for AI capabilities possible allowing any software to enhance any AI application.

It's All About the Idea ​

So we have AI with capabilities. How is it supposed to change the world?

Honestly, I can't answer, and I won't even try. We're talking about things that do not exist yet. Everything has to be imagined, and we're in that innovative phase where ideas have to be defined, tested, and analyzed.

The chance of failure is high, higher than the chance of success. But that's how innovation works. We are early in the technology adoption lifecycle.

We know it's possible, but "it" remains to be defined.

Chatting with Content ​

Two things made me realize that something had changed.

In March 2023, VS Code introduced a Chat view, but it wasn't working well and it wasn't able to edit files.

So in October 2023, I built a VS Code extension called Utils AI, allowing me to quickly run AI on my files. It was a way to quickly fix typos and improve SEO for my articles or documentation pages. It worked great.

It wasn't until November 2024 that VS Code introduced its edit mode with Copilot Edits, allowing the AI to modify files directly. You ask, it modifies intelligently. At least it tries.

At the same time, Google introduced NotebookLM, a new tool to interact with content, text, PDFs, videos, using natural language.

Since then, I discovered something new. You can literally chat with your content, having a conversation. Mind blowing.

Within VS Code, I asked Copilot to fix typos, then to suggest title improvements, add to bold important text, and tell me if I missed something or if a part was unclear. Clearly, my Utils AI extension became completely obsolete, but something even more powerful emerged.

Since then, I have an obsession in mind. How can I make something similar for my own website soubiran.dev? How can I chat to my content using natural language?

Recently, this tweet from Adam about Notion 3.0 made me realize the potential of a chat bar to interact with a website could go far beyond chatting. Keep it in mind; it will be part of another article.

Embracing Failure ​

I lost count of how many times I tried. However, I can confidently say that each was a failure, as you still can't chat to your content using natural language. But that's from failure that we learn the most.

1. In 2024, I tried to build a RAG using LangChain. I never created more than a small POC on my computer.
2. In February 2025, I built a search bar for my website. It's fully functional and works fine, but it's far from the experience I'm looking for.
3. Later, I tried to use a search engine like TypeSense. I also set up semantic search and natural language search. I even tried DocSearch to automate the process. All these solutions were good, but they were just search. Even the natural language search was just a search. Nothing was near a chat experience.
4. In April 2025, Cloudflare introduced AutoRAG: a fully managed RAG solution, from indexing to querying. I tried it; the overall system worked well, but it was dumb as a rock.

I failed again and again.

The Birth of an Idea ​

On October 1st, 2025, Benjamin tweeted about a new AI search for Nuxt UI. I instantly tried it, and it was exactly what I envisioned for my own website. Impressive step forward!

As soon as I get home, I started studying the code to understand how does he achieve such a result. I tried it again and again, I opened to Network tab of the Devtools to understand the request, and I deeply read the code to understand each word. Thank you the Nuxt team for making it open-source!

Something that instantly jumped to my mind was this idea of using their own MCP with an AI Agent to answer user queries. You could say it's overkill, I would answered that it's insanely powerful.

It was around 21h when I had an idea. What if I could create a similar system for my own website?

I started thinking about the solution, imagining problems, and trying to solve them all within my mind. Twenty minutes later, I had all the pieces needed to build something. I powered on my computer, opened OBS, launched a stream, and seven hours later the POC was working.

What a joy! I have no words to describe the feeling when you first see the AI using its tools and then perfectly answering user queries.

To be fair, the failures I encountered along the way were invaluable to achieving this so quickly.

An Opened Frame to the Future ​

It took me more than a year, nearly two, a lot of failure, to build a working prototype of my vision. Finally, it's here. Patience matters.

It's impressive and astonishing to see that level of precision in the answers. The AI is able to understand the subtext perfectly and provide relevant information.

The craziest part? It's just the beginning and I have so many ideas to explore to make the system even more powerful.

For now, it enables me to chat with my website, and this opens a new paradigm for the web. I call it:

Website as a Search Bar

This is a really thrilling concept. You can expect articles, technical deep dives, live streams, and even talks. You're a meetup organizer looking for speakers? Reach out! as it could be a talk.