Given that everybody right now is publishing AI agents on their websites, in the form of chats, my advice is: don’t do it.
What do I mean by that? There are two visions fighting each other in artificial intelligence: the agent as a browser, and the agent as a service.
The agent as a browser
Let’s take some examples. An AI agent as a browser means I have an agent that is more personal, more mine, closer to me. I query it, and it goes off to navigate databases, skills, MCP servers - to put it in practical terms. That agent then talks to my email, to Google Maps, to my management software, to the events website, and reports back what it finds.
In this sense the agent is a substitute for what we have always done with a browser: going to sites, searching, clicking, reading. It does all of that directly, so the mode of use is similar to agentic browsing - you ask for things, and it resolves them for you.
The agent as a service
The other perspective is the agent as a service. I land on a website and I find a whole chat there, an entire entity set up to solve my problem. This is more or less the vision of the agentic web: many agents exposed directly on the web, talking to each other. Not necessarily close to you, not necessarily your personal agent, but agents specialized in certain services - the Booking.com agent, the taxi agent, the agent for whatever comes to mind.
A lot of people are seeing it this way, so they work hard to put the entire chat online, the whole AI stack. And in my opinion that isn’t working in this historical moment, even though there are protocols like Google’s A2A, which is struggling to take off.
Users are attached to their own agent
What I see instead is a lot of attachment from the end user to their own agent, and from there they want to reach everything else. They don’t want to change sites or download applications in order to deal with an artificial intelligence.
It’s a subtle distinction, but I think it will make most of the difference in the coming period. Look at the success of MCP, the protocol that gives agents access to external services: bookings, statistics, videos - data, basically. The intelligence is yours, in your belly, next to you, and then through MCP that intelligence can navigate external resources. That’s the direction I think we’re going in.
So what do you sell?
In this historical moment, the winning strategy in my view is to build the services that sit around the artificial intelligence. Then, eventually, you can build the whole AI stack and put your agent on the web - but keep the services your agent accesses cleanly separated, so that you can sell those services on their own.
Because in my opinion the bulk of the market goes there: go for the services around artificial intelligence, not for the artificial intelligence directly.
If you want to go deeper into this, want more advice of this kind, and want to practice with agents - both building them and using them - write to me. That is my corporate course, and it’s called Reskill.
Vibe godo.
