Let Your Agents Drive the Browser (With You Watching)

It started with Iceland

I started changing my mind while planning a camper road trip in Iceland with five friends. I already wrote about planning that trip with AI, because honestly it was one of those moments where you realize this thing is not just for writing better emails or generating a list of restaurants. It can actually help you organize real life, with real constraints, real bookings, and real annoying websites.

At the beginning it was simple. Search places, compare options, look for activities, check dates, help me understand what makes sense. Fine. But then at some point you move from “find me something” to “book this thing”, and this is where everything becomes more interesting. Or more frustrating, depending on your mood and the website in front of you.

The moment you need to actually do something on the internet, the browser becomes the real interface. And the browser is messy. Every website has a different experience, a different checkout, sometimes a different language, sometimes a different idea of what a phone number should look like. For a human it is already annoying. For an agent, everything was frustrating.

The internet was not built for agents

I come from backend and API. So my instinct was always: give the agent an API, an MCP server, a CLI, something structured. If I need to configure something, I prefer a command line. If I need to access data, I prefer an endpoint. But sometimes you simply can’t: there is no API, only a website, a login and a form.

The funny part is that the internet was not designed to be consumed by agents, but agents are becoming some of the heaviest users of it. They search, compare, read, summarize, navigate, and now they start doing actions. You see all these discussions around personal-agent protocols and similar stuff, and the long story short is always the same: agents doing things on your behalf, including things that require authentication.

And yes, the moment credentials enter the conversation, people become nervous. Fair enough. I am not saying “give your whole life to a random agent and go to sleep”. In my case, they work in a dedicated Chrome profile, not my personal browser. I authenticate myself, then I let them continue from there, with me watching or checking depending on the task.

The browser became part of my workflow

This is the part that changed for me. More and more often, my workflow starts with: “open a new browser session, I will authenticate, then you do it.” It sounded very futuristic the first times. Also a bit wrong. Like, wait, am I really letting an agent move inside a logged-in service?

But then I looked at what was actually happening. It was not magic. It was not the agent stealing my laptop and deciding to change my life. It was me deciding the boundary, the speed, the level of autonomy, and the stop point. That is a very different thing.

You don’t give agents control once and forever. You decide how much control they have in that moment. You can say: fill this configuration step by step and stop before saving. Or you can say: do everything autonomously and come back when it is done. Or something in the middle, which is where most of the useful work happens.

I was discussing this with some colleagues, and the reaction is often: “No, I will never give my credentials to my agents.” And I understand the instinct. But my question is: why not, exactly? If you are the one logging in, if it is a dedicated browser profile, if you decide what the agent can do, and if you are watching the important steps, then the conversation becomes less dramatic and more practical.

Real examples, not theory

The Iceland planning gave me the first practical examples. The agents configured all my bookings without issues, and in some cases honestly better than me, because they were more patient with the stupid details. They filled the forms and moved through the flow without me having to babysit every single click.

At one point, a website did not accept an Italian phone number. The agent tried, understood the problem, and instead of getting stuck forever, it fell back to email. It sent them an email instead, the most boring but effective way possible.

Then I had another experience with Google Cloud configuration. I know AWS reasonably well, enough to move around, but I am not an AWS architect. With AWS, I usually install the AWS CLI, authenticate, and then the agent can do the majority of the work from there. I check things, especially if we are talking about DNS records or configuration that can create pain later, but the agent does a lot.

With Google Cloud, the experience became more browser-first. Open the browser, authenticate, let the agent configure, enable the services, return back to me. Same idea with some AWS forms too: it filled the questions, moved through the boring parts, and asked me when something actually required my judgment. Not everything needs my brain. Some things just need someone to read, click, and not get tired.

The fear is real, but also a bit outdated

I am not saying this is mature. It is not. If you do this today, you are still a bit of a pioneer. Sometimes the agent will be clumsy. Sometimes the website will be hostile. Sometimes you will stop it because you don’t trust the next click, and that is fine.

The control is not only technical. It is conversational. You see the action happening, you can stop it, you can slow it down. I can ask why. I can ask what changed. I can ask the agent to compare the current configuration with the expected one. I can ask it to pause before the dangerous part. This is not the same feeling as a black-box automation running somewhere in the background and hoping everything is fine.

Of course, use some consideration. Don’t start with your bank account or the most critical admin panel in your company. Start with low-risk tasks, a dedicated Chrome profile, and a clear instruction. Watch how it behaves. Then decide if it deserves more autonomy or not.

Top-down learning, again

For me this connects with something I already wrote about: top-down learning. Instead of studying a whole platform from the beginning, the agent does what I need to do, I control, I ask questions, and I learn from the real task. Not from a generic tutorial. From the actual configuration I needed today. If I am curious, I never stop learning.

So yes, I am changing my mind. I am letting agents drive the browser more than before, with me watching, and sometimes with me fully letting them do the boring work. It still feels early. It still feels a bit weird. But the more I try it, the more it feels obvious.

Maybe the next interface for agents is not a perfect API, not a new protocol, not a beautiful automation layer.

Maybe it is just the messy browser we already have, finally used by someone patient enough to click through it.