You ever feel like the browser is just a very expensive, very shiny window frame that we’ve been staring through for thirty years? We keep polishing the glass, but the view is always the same.
It is a bit of a relic, isn't it? We’ve moved from static documents to heavy web apps, but the basic paradigm of "point, click, and manage your own tabs" hasn't shifted since the nineties. Today’s prompt from Daniel is about the rise of the AI-native browser, and it basically argues that the window frame is finally being replaced by something with a brain. Think about the Netscape era—the browser was just a renderer for HTML. Then Chrome came along and made it an operating system for JavaScript. Now, we’re seeing the third act where the browser becomes a collaborator.
Right, he’s looking at things like Perplexity’s Comet, Arc Max, and Dia. And by the way, speaking of brains, today’s episode is actually powered by Google Gemini Three Flash. It’s writing the script while we sit back and pretend to be the ones doing the heavy lifting.
I’ll take the help! But honestly, this topic is fascinating because we’re seeing a split in how people think about the web. Is the browser a tool you use, or is it an agent you delegate to? When we talk about "AI-native," we aren't just talking about a sidebar with a chatbot pinned to it. We’re talking about browsers where the rendering engine, the tab logic, and the navigation are built around an LLM core. It’s like the difference between a house with a smart speaker in the corner and a house where the walls themselves are sentient.
It’s the difference between a car with a GPS glued to the dashboard and a self-driving Tesla. One helps you navigate; the other understands the road. But I want to push on that definition. "AI-native" feels like one of those marketing terms that could mean anything from "we have a shortcut to ChatGPT" to "we’ve redesigned Chromium from the ground up." What’s the actual technical threshold here, Herman? Is it just about the UI, or is there something deeper in the stack?
That’s the right question. To me, a truly AI-native browser has to do three things. First, it needs semantic understanding of the DOM—the Document Object Model. It shouldn't just see pixels or text; it should understand that "this button is for checkout" and "this text is a shipping policy." Second, it needs autonomous state management. It should be able to handle tabs not as a list of URLs, but as a structured context of what you’re currently working on. And third, it needs an action layer—the ability to interact with the web without you moving the mouse.
So, looking at the current crop, who’s actually hitting those marks? Daniel mentioned Comet from Perplexity. They’ve been the darlings of AI search, but a browser is a much bigger swing than a search engine.
Comet is interesting because it’s trying to solve the "tab overload" problem at the source. Launched in beta back in February, its whole pitch is that you shouldn't have to visit five different sites to compare information. It uses its internal LLM to synthesize pages in real-time. If you’re researching a product, it doesn't just give you a list of links; it actually browses those links in the background, extracts the specs, and presents them in a unified view. It’s essentially turning the browser into a real-time researcher. Imagine you’re looking for a new mountain bike. Instead of opening tabs for Trek, Specialized, and Giant, Comet reads all three simultaneously and builds a comparison table for you right in the address bar.
I’ve seen some of the early reviews. People are saying it feels like the browser is "reading ahead." But there’s a massive privacy elephant in the room there, isn't there? If the browser is "synthesizing" everything I look at, it’s basically a keylogger with a PhD. How do they handle encrypted sessions or sensitive data like medical portals?
You hit the nail on the head. If you look at the terms of service for some of these new players, it’s a bit of a wild west. They need to ingest your data to provide that "intelligence." If it’s summarizing your work emails or your banking dashboard to "help" you, that data is flowing through their models. It’s a huge trade-off between utility and total surveillance. Some are promising local inference—where the AI runs on your own chip—but for the really heavy lifting, it’s still going to the cloud.
And then you have Arc Max. The Browser Company has been trying to "fix" the browser for a while now. Their January update for Arc Max was all about context management. I think they claimed it reduced tab counts for power users by something like forty percent. I’ve tried it, and it does feel different, but is it "native" or just a very smart skin?
Arc Max is a great example of the "context manager" approach. Instead of focusing on browsing the web for you, it focuses on organizing your brain. It does things like "Tidy Tabs," where it uses AI to group your fifty open tabs into logical workspaces based on intent. Or "Arc Explore," where you type a query and it builds a custom "mini-site" for you by pulling from various sources. It treats the browser like a dynamic canvas rather than a file folder of websites. It’s less about replacing the website and more about rearranging the clutter so you can actually see the desk.
It’s the "Marie Kondo" of browsers. Does this spark joy? No? Okay, the AI will archive it for you. But let’s talk about the third one Daniel mentioned, because this feels like the most radical shift: Dia Browser. This isn't just about summarizing or organizing; it’s about agents. This is where we get into the territory of the browser actually "doing" things for us.
Dia is the one that really blurs the line between a browser and a robotic process automation tool. They released their agent SDK in March, and the goal is total delegation. You shouldn't have to go to a travel site, filter for "non-stop," select a seat, and enter your credit card. You should just tell the browser: "Book me the cheapest flight to Tokyo next Tuesday," and the browser—acting as a Dia agent—navigates the DOM, interacts with the scripts on the page, and fills the forms.
See, that’s where my skepticism kicks in. We’ve been promised "agents that do things" for a year now, and mostly they just get stuck in a loop trying to click a "cookies" pop-up. I remember trying an early web agent that was supposed to order pizza, and it spent ten minutes trying to "agree" to a newsletter it couldn't close. What makes an in-browser agent different from, say, a specialized LLM wrapper?
The difference is direct access to the browser’s internal state. Most "wrappers" have to look at a screenshot of the page or try to parse a messy HTML dump. An in-browser agent like Dia lives inside the rendering loop. It sees the accessibility tree, it understands the JavaScript execution, and it can simulate events at a much more granular level. It’s not "looking" at the screen; it’s part of the engine. That makes it significantly more robust than a bot that’s just trying to guess where the "submit" button is based on a picture. It can see that a button is disabled because a specific validation script hasn't run yet, whereas a screenshot-based AI would just keep clicking it fruitlessly.
Okay, but if the browser is doing the clicking, what happens to the web as we know it? If I’m a web developer, I’ve spent twenty years designing for humans. I want you to see my beautiful CSS, my clever ads, and my "sign up for our newsletter" pop-up. If Dia just bypasses all that to scrape the data or click the button, I’m losing my mind—and my revenue.
This is the "Agentic Internet" problem we’ve touched on before, but it’s hitting a boiling point with these browsers. We might be heading toward a "Clean Web" protocol where sites provide a machine-readable version of their interface specifically for these agents. Because right now, there’s a literal arms race. Sites are using bot-detection to block agents, and browsers are using AI to look more like "human" clickers. It’s a mess. Imagine a world where a website detects an AI agent and serves it a completely different, high-speed text interface while the human gets the pretty, ad-heavy version.
It’s hilarious, really. We’re using the most advanced technology in human history to trick a website into thinking a sloth is actually clicking a button to buy socks. But Daniel asked a really pointed question: if we have these in-browser agents, do we still need things like Playwright or Anthropic’s "Computer Use" feature?
I think they serve fundamentally different masters. Playwright and Puppeteer are developer tools built for scale and reliability. If I’m a developer and I need to scrape ten thousand product pages every hour for a price comparison site, I’m not going to open a Dia browser window and watch an agent click around. I’m going to use Playwright in a headless environment on a server. That’s batch automation. It’s about "X leads to Y" in a repeatable, programmatic way.
So Playwright is the factory assembly line, and the AI-native browser is more like... a personal assistant sitting at a desk?
Precisely. The AI-native browser is for "one-off" high-complexity tasks that require real-time reasoning. For example, if I ask an agent to "find a restaurant in Tel Aviv that has outdoor seating, is open on Saturday, and takes reservations through a specific app," that’s a reasoning task. The agent has to navigate, read reviews, check the "about" page, and then interact with a reservation widget. You could code that in Playwright, but it would take you hours to account for all the edge cases. An AI agent does it by "thinking" through the UI. It can handle a pop-up it’s never seen before because it understands the intent of the pop-up.
And what about Anthropic’s "Computer Use"? That’s the one where Claude literally takes over your mouse and keyboard and looks at screenshots. That feels like a "brute force" version of what Dia is doing.
It is. Anthropic’s approach is "general purpose." It can use Excel, it can use Slack, it can use the browser. But because it relies on visual interpretation—literally looking at pixels—it’s prone to "hallucinating" a button that isn't there or getting confused by a loading spinner. A browser-native agent has a "cheating" advantage. It doesn't need to look at the pixels to know the button is there; it can see the code. In terms of reliability for web tasks, the browser-native approach will almost always win. It's like the difference between someone trying to drive a car by looking through a camera feed versus the car's internal computer knowing exactly where the wheels are positioned.
It feels like we’re seeing a convergence, though. Eventually, the operating system and the browser become the same thing for most people. If my browser is where I do ninety percent of my work, and my browser has a built-in agent, do I even care about the rest of the OS?
That’s the bet Perplexity is making with Comet. They want to be the new OS. If they own the browser, they own the context. They know what you’re researching, what you’re buying, and what you’re writing. That is a massive amount of leverage. Google knows this, which is why they’ve been frantically layering Gemini into Chrome. But Chrome has "legacy baggage." It has to support billions of users and a massive ad ecosystem. Startups like Dia or The Browser Company can move faster because they don't care if they break the "ad-supported" model of the web.
Speaking of Google, I find it funny that they’re the ones lagging here. They literally invented the transformer architecture that makes all this possible, and yet Chrome still feels like a browser from 2018 with a "Help me write" button tacked onto the text fields. Why is the giant so slow to move on its own turf?
It’s the classic Innovator’s Dilemma. If Google makes Chrome "too smart," it might stop people from clicking on search ads. If an agent just "gets the answer" or "books the flight," the user never sees the search results page. They’re cannibalizing their own revenue. Perplexity doesn't have that problem. They want to kill the search results page. They want the browser to be the final destination, not the starting line.
Let’s look at some other players. We’ve talked about the big three, but are there any "dark horses" in the AI-native browser space? I’ve heard whispers about some more niche projects.
There’s a project called Atlas that’s been making waves. It’s focused specifically on "spatial" browsing. Instead of tabs, it treats the web like a giant zoomable map where AI helps you lay out connections between different sites. Imagine a canvas where your bank statement is next to your budget spreadsheet, and the AI is drawing arrows between them to show you where your money went. And then there’s SigmaOS, which is very popular in the developer community. They’ve integrated AI into their command palette, so you can basically "code" your browser behavior on the fly. You can hit a shortcut and say "summarize all open tabs about React hooks into a markdown file," and it just does it.
I’m still waiting for the "Sloth Browser" that just summarizes everything into three bullet points and then automatically hits "snooze" on my notifications. But seriously, for someone listening who is a developer or a power user, how should they be thinking about this? Is it time to ditch Chrome and move into a beta browser? Or is the friction of moving your passwords and bookmarks still too high?
I think for developers, the real takeaway is to look at the SDKs. Don't just look at these as browsers; look at them as platforms. If you can build a Dia agent that handles a specific workflow for your company—say, "update the internal CRM based on these three LinkedIn profiles"—you’re saving yourself a ton of custom integration work. Instead of fighting with APIs that might not exist, you’re using the UI as the API. It’s a paradigm shift in how we think about software interoperability.
"The UI is the API." That’s a great line. It’s also a terrifying line if you’re a UI designer. Your "user" might soon be a model running on a server in San Francisco rather than a person with a mouse. Does that mean we stop caring about how things look and only care about how they're labeled in the code?
It changes everything about "accessibility," too. We’ve spent years pushing for ARIA labels and semantic HTML for screen readers. It turns out, those same standards are what make AI agents successful. If your site is accessible to a blind user, it’s probably accessible to an AI agent. In a weird way, AI might finally force the web to be properly structured. Developers who ignored accessibility for years are going to start caring about it very quickly when they realize it's the only way to make their site "agent-friendly."
That’s a silver lining I didn't expect. "AI: The ultimate enforcer of web standards." But what about the "agentic" part of the internet Daniel mentioned? If everything is an agent, do we need a new protocol? Like, instead of HTML, do we need "Agent-ML"?
There’s actually work being done at the W3C right now on "Agentic Protocols." The idea is that a website could broadcast its capabilities directly to an agent. Instead of the agent having to "guess" how to book a flight by clicking buttons, the site says, "Here is my booking endpoint, here are the required fields, and here is the price." It’s basically a dynamic API that describes itself. We're moving toward a world where the "view source" of a page isn't just code, but a set of instructions for a machine to understand the business logic of the page.
So we’re going from a web of pages to a web of services, and the browser is the negotiator. I can see that. But I keep coming back to the performance. Running an LLM inside a browser—or even hitting an API for every "thought" the browser has—that’s got to be slow. Arc Max feels snappy, but is it doing real reasoning? Or is it just clever pattern matching?
That’s where the "Small Language Model" or SLM trend comes in. A lot of these browsers aren't running GPT-4 for every tab. They’re running much smaller, highly optimized models locally on your machine—using your GPU or the NPU in your new laptop—to handle the basic stuff like summarization and tab grouping. They only "call home" to the big models when you ask for something complex, like "plan a three-week trip to Italy." This hybrid approach is key. You don't need a supercomputer to rename a tab; you just need a model that understands the difference between "YouTube" and "Productivity."
Which brings us back to our sponsor, Modal. If you’re building one of these browsers or an agentic platform, you need that bursty, serverless GPU power to handle those "big brain" moments without paying for a cluster that’s idling twenty-three hours a day.
The infrastructure behind this is just as important as the UI. If the "agent" takes thirty seconds to think about every click, no one is going to use it. You need sub-second inference to make it feel like the browser is actually "native" and not just a slow-motion robot. We're seeing a lot of innovation in how these weights are quantized and served so that the browser feels like a living thing rather than a lagging utility.
So, let’s get practical. If I’m a listener and I want to "live in the future" for a week, what’s the move? What's the specific workflow I should try to offload to an AI browser?
I’d say download Arc Max first. It’s the most "polished" version of this vision. Use it for a week and see if the "Tidy Tabs" and the AI-generated summaries actually change how you work. Try their "Ask on Page" feature where you can hit Command-F and instead of searching for a word, you ask a question like "Does this paper mention the sample size?" Then, if you’re feeling adventurous, get on the waitlist for Perplexity’s Comet. That’s the "search-first" vision of the web, and it’s a very different experience. It’s less about "visiting sites" and more about "consuming information."
And if you’re a dev, play with the Dia SDK. See if you can automate that one annoying task you have to do every Monday morning—the one that doesn't have an API. Maybe it's checking for broken links on a staging server or pulling data from a legacy corporate portal. That’s where the "aha" moment usually happens. When you see the browser navigate a complex, legacy web portal on its own, it feels like magic.
It really does. But we have to be realistic about the "agentic" future. We’re in the "awkward teenage years" of this tech. It’s going to make mistakes. It’s going to accidentally buy the wrong flight, or it’s going to hallucinate a "delete account" button when you asked it to "change password." We’ve seen examples where agents get stuck in "infinite scrolling" loops because they think they haven't found the answer yet. We’re still in the "trust but verify" phase. You wouldn't give your car keys to a teenager who just got their permit; you shouldn't give your primary credit card to an experimental browser agent without limits.
I don't even trust you to pick the right restaurant when we’re out, Herman. I’m definitely not trusting a browser agent to book my flights yet. But I love the idea of the "tab manager." My browser currently looks like a graveyard of good intentions. Fifty tabs open, and I have no idea why half of them are there. If an AI can just look at my open tabs and say, "Hey, these twelve are related to that vacation you were planning, can I group them?" that's a huge win.
And that’s the low-hanging fruit. If AI can just solve the "mental load" of managing the browser, it’s already won. The "autonomous agent" stuff is the moonshot, but the "smart organizer" is the immediate value. Think about how much time we spend just looking for the right tab. If the browser knows your intent, it can surface the right tab at the right time.
What’s the "endgame" here? Does Google just eventually buy one of these startups and fold the tech into Chrome, or do we see a genuine shift where Chrome becomes the "Internet Explorer" of the 2020s—the browser you only use to download a better browser?
I think Google is too big to fail in the short term, but the "default" is definitely under threat. If OpenAI releases their rumored browser—which would presumably be built around "SearchGPT" and their latest models—that is a direct attack on the heart of Google’s business. If your browser is "ChatGPT with a URL bar," why would you ever go back to a standard browser? We could see a world where the browser market fragments again, not based on rendering speed, but on the "personality" and "capability" of the built-in assistant.
It’s a battle for the "front door" of the internet. For twenty years, the front door was a search box. Now, the front door is an agent. And that agent is going to be sitting right inside the browser chrome, watching everything we do.
And that agent needs a home. The browser is the natural habitat for an AI agent because that’s where all the data is. It’s where our identities are, our cookies, our history, our work. Moving the agent "outside" the browser—like Anthropic’s computer use—is a clever hack, but moving the agent "inside" the engine is the long-term architectural winner. It's about reducing the friction between the thought and the action. If the agent is part of the browser, there's zero latency between "I want to do this" and the browser executing the code.
It’s like the difference between a ghost haunting a house and the house itself being haunted. I’d rather have the smart house. I want the house to know when I'm hungry, not just have a ghost that tries to move the spoons around.
Or a smart sloth.
Hey, I’m plenty smart. I just move at a pace that allows for deep reflection. Which is exactly what we need when the web is moving this fast. It’s easy to get caught up in the hype of "everything is changing," but the fundamental human need—to find information and get things done—is the same. The "how" is just getting a lot more interesting. We're moving away from the era of "manual browsing" into an era of "curated navigation."
I think we’ve covered the "how" pretty extensively today. From the semantic DOM understanding of Dia to the context management of Arc Max, the "AI-native browser" isn't just a trend. It’s the next logical step in the evolution of computing. We’re moving from the "Personal Computer" to the "Personal Agent." It's the fulfillment of the vision people had in the 60s—the "computer as a colleague" rather than a typewriter.
And on that note, I think it’s time to let the agents take over. If you want to dive deeper into any of these, we’ll have links in the show notes at myweirdprompts dot com. We'll include the documentation for the Dia SDK and some of the white papers on semantic DOM parsing if you really want to get into the weeds.
We should also mention that if you’re interested in the "agentic internet" side of things, we did a deep dive into Google’s MCP standard a while back—that’s the Model Context Protocol—which is a big part of how these agents will actually talk to the web. It’s worth a look if you want the "plumbing" side of this conversation. Understanding how the data flows from the site to the model is going to be a crucial skill for devs in the next few years.
Look at you, making a helpful suggestion. You’re basically an AI agent yourself at this point, Herman. You've been fine-tuned on too many technical manuals.
I’ll take that as a compliment. At least I don't hallucinate as often as the early betas of Comet.
Take it however you want, as long as you take us to the credits. I've got fifteen tabs about artisan coffee to organize, and I'm hoping my browser will do it for me.
Fair enough. Huge thanks to our producer, Hilbert Flumingtop, for keeping the gears turning behind the scenes and making sure our agents don't go rogue during the recording. And a big thanks to Modal for providing the GPU credits that power our research and this very production.
This has been My Weird Prompts. If you’re enjoying the show, do us a favor and leave a review on Apple Podcasts or Spotify. It’s the only way the algorithms know we’re worth talking to. If the AI is going to take over the world, we might as well make sure it likes our podcast.
We’re also on Telegram if you want to get notified the second a new episode drops. Just search for My Weird Prompts. We share a lot of the raw prompts and research notes there that don't make it into the final cut.
All right, I’m off to see if I can get a Dia agent to write my emails for the rest of the day. Maybe I'll finally hit Inbox Zero.
Good luck with that. Just make sure it doesn't accidentally resign from the show on your behalf. See you next time.
See ya.
