Web Browsers, Agentic AI, and the Future (Premium)

I can’t stop thinking about web browsers. Smaller web browser makers are starting to innovate in unique ways ahead of bigger changes that might occur thanks to AI. This work is overdue, I think. It’s not like we nailed the web browser UI in the 1990s and then never needed to make significant changes to it over the intervening 25+ years, despite all the changes happening around these apps in personal computing during this time. But the future is elusive. It’s not clear what’s next.

What is clear is the past. I made the case in my Programming Windows series, which became the book Windows Everywhere, that the decline of Windows can be tied directly to Microsoft’s mistake in reversing its decision to fully embrace the web. And it is not coincidental that Apple today is taking a similar protectionist stand with iOS and the iPhone in trying to kneecap not just competing web browsers, but also the cross-platform app platform and extensibility they would otherwise support.

But web browsers will not be ignored. They are the most popular and most often used app across desktop and mobile. They are now a conduit for AI-based services. And they are the one app, the one interface, that works everywhere and is common to some degree across platforms. Today, we seem poised on a precipice of long overdue change.

Sort of. Steven Wright told a joke in the 1980s that perfectly encompasses how I view this situation today.

“You know when you’re sitting on a chair and you lean back so you’re just on two legs. and then you lean too far so you almost fall over, but at the last second you catch yourself? I feel like that all the time.”

Yes. That.

So where is this all heading?

? Step into our agentic future

Microsoft CEO Satya Nadella once described Copilot as being “like the Start button” in Windows, in that it would “become the orchestrator of all your app experiences.” We naturally associated Copilot specifically with Windows at that time, as this was when Microsoft announced it was adding a Copilot app to Windows 11, after all. But the ways we’ll use AIs like Copilot extend far beyond Windows apps and PCs. More to the point, Nadella was purposefully comparing this future interface to the app launching system for which Microsoft is still most famous. But the “apps” that Copilot will orchestrate will more commonly be what I think of as online services, and Copilot as an interface is something that needs to be available everywhere, not just on PCs.

More recently, Microsoft has described “the second wave” of Copilot advances as agentic, meaning that its AI can be configured to work autonomously on your behalf to complete tasks. And not just tasks, but complex multistep tasks that might require it to interact with multiple online services and apps.

This is orchestration writ large. At a high level, an operating system on a PC, server, or mobile device abstracts the hardware for developers, and it provides an interface for interacting with the apps and services it provides in software. In both cases, the OS provides orchestration capabilities: You tell it what you want and the OS handles the low-level details. AI will do the same thing, but more broadly, by abstracting how it does what you ask it to do. At first in a ham-handed way–like asking Copilot in Windows 11 how to enable dark mode, an act that takes more time than just Googling it as I write this–and then in ever more sophisticated ways.

Nadella’s late 2023 comparison of Copilot to the Windows Start button seemed a bit simplistic and even silly at the time–few could imagine that Copilot would be how we’d “run apps” or whatever at the time–but perhaps he was onto something. Windows is semi-unique in personal computing OSes in that it does, for the most part, offer a single place for users to go to initiate whatever interaction (Start, rather than a grid of icons you configure on mobile). It has even had integrated search capabilities for decades–once literally called Start Search–an interface by which you tap a button and just start typing.

It’s interesting how that parallels how we interact with Copilot and other AIs today. Coincidental, yes. But interesting.

?️ What’s old is new again

This interaction–type (or say) something to get something back–isn’t unique to Windows and OSes. We interact with the web this way, too. We click the address bar in a web browser, we type a search term, and search results start appearing as we type. Again, this is type/say something to get something. And when we get something we like, want, or need, we can save it as a bookmark or perhaps share it to another service, for example a read later service, or, more commonly, do something. That is the reason we searched in the first place.

Agentic AI promises to automate this process.

Automation in computer systems has rarely gone well. And while this is perhaps unfair, it’s particularly true in the Microsoft space. We all remember the goofy Office Assistant, Clippy, that would pop up as we started writing in Microsoft Word, asking us if we wanted help. Clippy–which Bill Gates hated and referred to as, “that f#$king clown”–was well-intentioned, but like so many things, it had unintended consequences and was reviled by most users. It was disabled by so many customers that Microsoft gave up on it after two product versions.

But Microsoft never really gave up on the ideas that led to Clippy. Microsoft doesn’t get a lot of credit for this kind of thing, but it kept trying, again and again, to be helpful.

Cortana, the personal digital assistant that Microsoft introduced first on Windows Phone in 2014, combined a voice interface with personalization and a hopeful helpfulness. Microsoft Band, also from 2014, was a sensor-laden wearable that presaged today’s smartwatches and promised to combine the personal and health information it gleaned about you with your Outlook schedule to deliver holistic improvement advice. Even something like Focused Inbox, which has polluted every version of Outlook imaginable with its automatic email sorting, is an example of this kind of well-intentioned but mostly unwanted help. This is Microsoft trying, again and again, to be helpful but always just getting in the way.

Can Microsoft do with agentic AI what it could seemingly never do before in almost any of its products and services? I can’t answer that question, of course, no one can, but hope springs eternal. And as with the conversational Siri problems that Apple is now experiencing, it almost doesn’t matter. Whether it comes from OpenAI, Google, Anthropic, and any other companies, this capability is rolling out as I write this, and across the industry. A better question, perhaps, is where we will typically use agentic AI. If today is any guide, and I think it is, that, too, is complicated. And it’s complicated because we’ll likely use it everywhere. It’s AI, after all. It will be everywhere.

If that’s true, and I think it is, then the overall dynamic of how we use and interact with personal technology won’t change. We’ll still use word processors and note-taking apps to write. We’ll still use spreadsheets to crunch numbers and make pretty charts. We’ll still spend more of our time interacting with technology on smartphones than we do on any other devices, though we will also all still be multi-device users, and will interact with technology, and more specifically AI, everywhere. And the app we will use the most, almost certainly, will still be a web browser.

The caveat here is that browsing may be the thing we lose over time, like a vestigial capability that a coming generation forgets and then has to be reminded why these things are called browsers in the first place via whatever passes as a meme in this future imaginary world. OK, that was a mouthful. What I mean is that if agentic AI works, and it will eventually work just fine, then we will be doing less browsing and more doing. Or, AI will handle the doing and we will simply tell it what to do.

Surprise. We’re already doing this.

?️ What’s new is old again

The canonical example of this type of functionality today is a price-checking service. For example, I use (and pay for, go figure) a service called Going to monitor the price of flights between Newark, New Jersey and Mexico City. Going emails me regularly, sometimes about particularly good flight prices to anywhere, and sometimes about the specific flights I asked for. And when I see a deal that matters, I forward it to my wife or discuss it with her. Many others get alerts from other price-related services. For example, you might monitor the price of an Xbox, phone, or other electronics item, either broadly or at a specific store like Amazon, and then be alerted when there’s a sale. (A simpler example, perhaps, is an RSS reader, which brings the notion of push to news reading. With RSS, I don’t browse for news stories, they come to me.)

This type of thing is common because it’s inherently useful. But it could be even more useful.

Google, The Browser Company, Opera, and others have all signaled or even previewed the work they’re doing to automate this process fully using AI. And in this agentic future, one could optionally instruct an AI not only to look for a specific flight deal, as I am currently doing, but to book a flight when certain conditions–date ranges, time frames, cost, and so on–are met. That is, instead of getting an email from Going, I might in the future be working on a laptop or tablet or whatever and get a pop-up notification about a particularly good deal and be prompted to make the purchase. I will have already configured it to know which credit card to use, and within seconds, I could suddenly find myself with a newly booked trip and I barely lifted a finger.

This description is obvious enough, I know. But so, too, is the reaction from certain quarters. There’s a Ven diagram out there in which the overlapping audience between a circle representing those with privacy concerns and a circle representing AI deniers is right now writing me frustrated feedback, arguing that the fantasy I’ve concocted is either impossible or at least prone to mistakes. They collect stories about AI fails, like the Apple Intelligence news summary notification no one can seem to forget or forgive, like totems, proof points of their biased worldviews.

But my mind often drifts to contrary examples. For example, the Apple community loves to share every story they can find about an Apple Watch that somehow saved someone’s life. But what never gets reported–and this has to be far more common–is when someone dies while wearing an Apple Watch. And I feel like with AI, the failures, especially in the early days of this fast-moving era, are notable in whatever ways while also being increasingly the exceptions. Like shark attacks, big news, yes, but also rare. No matter how many times I mention how incredible AI is with software coding, for example, I always get at least one piece of feedback pointing to an example where it failed them.

The important thing to remember about AI, however, is the speed. Experiences with AI-based services that didn’t work well six or 12 months ago are no longer relevant. And while many have a bad experience and then give up, that kind of feedback loop doesn’t work in the AI era. The improvements are real, and they really do happen very quickly, whether you experience them or not. I’ve written before that we collectively need to get past AI denial, that that ship has sailed, and that a more pragmatic approach is required. We each need to think about, research, and then implement AI where it can help us the most. And these evolving agentic capabilities are perhaps the best example of how AI can help save us time, now and even more so in the near future.

? Back to the future

But let’s return to web browsers, the logical, mainstream-friendly interface for agentic AI. At some point, we’re all going to have multiple agents working on our behalf, popping up from time-to-time with updates, either asking us questions, looking for permission to move forward, or alerting us to something important. This future does raise questions in my mind, of course. But they’re not really tied to privacy or accuracy, two issues that I feel will solve themselves.

The first problem is one that pre-dates the modern AI era and something of a personal concern. Thanks to what is now a fire hose of information that arrives all day long every day, many of us can no longer focus or consume long-form content easily. For me, the biggest victim here is reading. I love to read, I read every day, and I am, of course, a writer. But when I was younger, I could fly through very long books quickly. And these days I cannot. It’s a problem, and if agentic AI means I’m doing less browsing–which, to me, often means reading things online–then it may only exacerbate this issue.

The second problem is related to time. If AI is saving me so much time, as I think it will, then I need to figure out what I’ll do with that time. This is another one of those contrary examples, perhaps a weird ADHD-related brain wiring thing. But where most AI critics will opine about job losses and other negative impacts, I find myself wondering what a future in which drudgery is removed from life bit by bit even looks like. There are science fiction depictions of this type of future–the fat humans gliding around on floating disks, unable to physically do anything in WALL-E, for example, inspired no doubt by one of Isaac Asimov’s Robot stories–that are meant to be troubling and inspire some introspection. Which they do, of course.

Perhaps I could make lemonade out of these problems, so to speak, and use that spare time to read more. But I kid, and I suspect that it’s only a matter of time before I’m using AI to summarize long form articles I can’t be bothered to read on my own. Like most people, I’m inherently lazy. But I also don’t want to end up on one of those floating disks.

Joking aside, it’s fair to say that this future will be whatever we make of it. And while nothing is certain, I do think the net outcome here will be positive. That’s the typical outcome of any technology transition like this. Damn the unintended consequences.