My first month building AI browser

The browser war has begun.

On June 11, 2025, The Browser Company released Dia in beta: an AI browser with many of the same ideas we had been exploring, but with more polish. Shortly after I joined the team, on July 9, Perplexity launched Comet for Perplexity Max subscribers. There are other smaller players in the space too, but it feels like only a matter of time before every major AI company starts treating the browser as strategic territory.

After one month of working on an AI browser, here are the questions I keep coming back to.

Handing control back to users

The search box used to have a relatively simple job: help people navigate or search. Now it has to do more.

It needs to support chat. It needs to understand more context. And it needs to communicate all of that clearly without making a familiar behavior feel suddenly complicated.

Our first version of Magic Box was built around the idea that the browser could infer the user’s intent: whether they wanted to navigate, search, or chat. In theory, that felt magical. In practice, it often wasn’t.

The problem is that intent is hard to guess from a single keyword, or even a short phrase. When the browser guessed wrong, users didn’t experience it as intelligence. They experienced it as friction. The “magic” stopped feeling magical.

The conclusion we came to was pretty simple: for now, it is better to preserve the behavior people already understand, then layer new AI capabilities on top of it. Users should feel like they are gaining power, not losing control.

Tab clutter is still the browser’s biggest pain point

After talking to 30+ people, the most common browser pain point was not surprising: tabs.

For years, browsers have treated tab management as a user-organization problem. We have tab groups, vertical tabs, pinned tabs, spaces, and workspaces. All of them help, but they still assume that the user wants to manage the mess manually.

One approach we have explored is automatically grouping tabs. Some users loved this. Others really disliked it.

Another polarizing idea was auto-archive: tabs that have been inactive for a certain period of time get moved out of the active workspace. My read is that this is less about the feature itself and more about discoverability, trust, and communication. Auto-archiving is not an entirely new idea, but users need to understand what happened, why it happened, and how to undo it.

The more I think about it, the more I feel like the tab problem needs to be reframed. Maybe it is not really about reorganizing tabs better. Maybe the better question is: what are people trying to accomplish with all these tabs?

A messy tab bar may look like clutter, but it often represents something in progress: a decision, a task, a trip, a purchase, a thought, a half-finished rabbit hole. It is scattered and unorganized, but there is usually intent hiding inside it.

That feels like the real opportunity for AI browsers.

Maybe the AI browser gap isn’t the biggest hurdle

Before starting this work, my biggest hypothesis was that people had a major pain point moving between AI tools and the browser.

That was certainly (somewhat) true for me. If I wanted help from AI, I had to copy and paste URLs, explain what I was looking at, give context, and then ask the actual question. It felt obvious that the browser and AI should be closer together. Wouldn’t it be nice if I could skip all of that?

This is probably why the AI side panel became such a common pattern. The browser already knows what page I’m on, so the assistant can help me with the context in front of me.

But in user interviews, more than a few people said they found less value in that connection than they expected. When they are browsing, they often do not want that much help. They are scanning, clicking, wandering, comparing, or just looking around.

When they do want AI help, it tends to be more specific and goal-oriented. They are not simply “browsing” anymore. They are trying to make a decision, understand something, summarize a set of information, complete a task, or move a workflow forward.

That distinction matters. The value of AI in the browser may not be “help me with whatever page I’m on.” It may be “help me turn this messy browsing activity into something useful.”

Proactive is the big bet

There has been a big push to make the browser a proactive assistant. I agree with the ambition. A personal assistant sounds great.

But I want a smart assistant, not a 15-year-old intern guessing what I’m doing.

That is the hard part. There is a gap between what people are doing in the browser and what they are ultimately trying to achieve. Today, much of that intent is implicit. It lives in open tabs, repeated searches, half-read articles, abandoned carts, saved links, and unfinished tasks.

We still need to understand what that intent looks like in an AI-native world.

The harder part may come after that: gradually reshaping people’s expectations of what a browser can do. People do not automatically expect their browser to help them think, decide, remember, or act. That shift will probably need to happen slowly, through moments that feel genuinely useful rather than overly eager.

What does that ultimately look like?

I’m not sure yet. I’m only a month into the journey.

But I do think the future browser will be less about pages, and more about the work people are trying to do across them.