I Stopped Browsing. My AI Agent Didn't.
How agents are replacing the web. (And apps, too.) A personal journey.
For 30 years, the web was a place you went — and the browser was what took you there. It’s the original user agent: You point it at the world, it goes and fetches, you read what comes back. I’ve spent my career on that little piece of software that acts on your behalf (I’m now Mozilla’s CTO), so it’s strange to admit that I mostly don’t drive it myself anymore. I point Zora, the name I gave my Hermes agent, at the world instead, and she goes. Browsing isn’t dead, but the job the browser has done for decades is quietly moving to something new, and I’m living that move. (Disclosure: Mozilla Ventures is an investor in Nous Research, the company behind Hermes. None of this is Hermes-specific, though — it’s the shape, an agent you can hand a skill and a cron.)
I can see it in my own logs: My internet usage looks nothing like it did a year ago. (I instrumented my Firefox to be sure.) I spend my hours in a chat window, not out on the open web. That’s my consumption falling, and I’m not alone. AI Overviews now reach more than 2.5 billion people a month, most Google searches end without a click, and when an AI answer shows up, people click out to the open web about half as often. Why is the web being read less? Because something else does the reading now.
My actions are down, too. I don’t open apps, I don’t fill in forms, I don’t click through three screens to book something. Or at least I’ve stopped doing it myself. Now, I hand it to Zora. Even the way I reach her changed: an action is now a Slack message or an iMessage. “Do a thing” is now just a text.
How it started: I wrote a sentence, and the harness did the rest. Zora can write code, run it against the real thing, watch it break, and rewrite, so a sentence becomes a script becomes a standing tool that just runs. She’d hit a wall, I’d catch where she was wrong, and we’d fix it together until it stopped being wrong.
This isn’t a story about reading less. It’s about handing over the doing, and the three things that are moving at once underneath it. I’ll keep coming back to all three, because together, they’re the whole point.
The thing that acts for me has moved up a layer. For 30 years, that thing was the browser; now it’s the agent. I send Zora instead.
The software didn’t get replaced by better software — it melted. Every job Zora took over used to be an app, or a form in it, and what replaced it was nothing you could download.
(The one underneath the other two.) Whoever holds the agent holds the keys to everything it reaches, everything it remembers, everywhere it goes wearing my name.
They’re small when it’s just me and my lunch order. They get a lot bigger when it’s your customers, your product, and your entire company — and that’s where this is headed. But the proof is personal, so that’s where I’ll start.
Before we get to the three things I’ve handed Zora, let’s look at the setup that makes it possible. Even if the gear sounds like a lot, stay with me, because the punchline is that you need almost none of it.
First, the Rig
People often ask about my hardware, so let me get it out of the way. (And then I’ll tell you how little of it matters…)
There’s more than one box. I already had a machine in my closet, running Ubuntu. Zora lives in a Docker container on it. A Mac mini sits beside it to reach into the Apple world — BlueBubbles for iMessage, Apple Notes, Contacts. Because I like to tinker with local models, there’s also a DGX Spark on the network and an old Bitcoin-mining rig with a row of RTX 4090s bolted to the 80/20. I’m using GLM-5.2 to do most of the actual thinking these days.
Here’s the honest part: That rig is what let me do all of this, but you can do most of it with a sliver. The one thing you truly need is an always-on box running an agent with a memory and honest access to your services, and that can be a $300 mini-PC or a cloud container. The Mac mini is just for deep Apple reach. The DGX Spark is just so a model can run on my own metal, where my data never crosses my own firewall. I could swap GLM for an equivalent tomorrow and not a single thing would change.
The model is the current in the wires — interchangeable and, honestly, a little boring now. The wiring is what makes it a house: a memory of who I am, a body that never sleeps, and my keys to the doors. I keep those keys in my pocket; the whole rig is just an elaborate way of making sure they’re mine, and it’s where everything here finally lands.
So let me get specific. You don’t need another essay telling you people are using agents now; you need to see exactly what that means on one machine: mine. Here are three things I actually hand Zora, hood open. Think of them as the ground floor: Once you can see precisely what’s already possible, the rest of the house is easier to see (and build).
#1 Everything I ate today…
Like a lot of people, I’m trying to keep my weight in check, which means logging what I eat. And logging meals is miserable. You finish lunch, open an app, search a database for “side salad,” guess whether it’s one serving or two, tap through a couple of screens, then do it all again for the slice of pizza.
So now I just text Zora: “had a slice and a side salad from State of Mind Pizza in Palo Alto.” That’s the whole interaction. By the time my fork’s down, it’s in MyFitnessPal — calories, protein, carbs, fat — with a window to undo it if the guess was off. (Snap a photo and it works the same way.) I didn’t search. I didn’t open the app. In fact, the only thing I open MyFitnessPal for now is to look at its charts. Everything else runs through Zora. She does the writing — and the reading, too. Ask what caloric budget I’ve got left for the day and she pulls it live. The app has quietly become a database I never sign into.
The wild part: MyFitnessPal has an API, but it’s shut to new developers. There’s no official door I could get Zora a key to. (I’d gladly have used the front door; an open API is the kind of thing I spend my days arguing for.) I didn’t solve it; Zora did. I never touched the auth, never looked at a header or an endpoint. I typed one sentence: “I want to start logging my food into MyFitnessPal,” and watched her work it out. At one point, she asked me to log in so she could grab the cookies, and then she let herself in the same side way the app’s own website uses.
Then we debugged it together. She’d ship a version, I’d catch where it was wrong, she’d fix it. The session cookies expire every few minutes, so she swapped them for an auto-rotating token and stopped fighting them. An entry once logged with its carbs quietly set to zero — the app shows a missing number as zero, so it looked right — so she added a check that refuses an incomplete entry rather than write a plausible lie. Photos kept undercounting pickled and fried foods, so she learned to look up the real numbers for such dishes first. I specified none of this. We found it by living with the thing, and she wrote the fix each time while I watched.
I asked for something the app was built to not allow, and a piece of software found its own way in and then corrected its own mistakes. A year ago, that sentence would have ended at “the door’s locked, sorry.” It’s the thrill of vibe coding, except it never closes into a session — it just runs in the background of my life. And even though I vibe-code all day now, I’m still a little stunned by how little I had to hand over to start.
#2 My calendar(s)
“What’s on my calendar?” I ask. It turns out I have a lot of calendars — family, personal, work — and usually an answer lives across all three. Zora doesn’t cram them into one view; she holds all of them in her head at once, so I can ask things like: Am I free Thursday afternoon? What does next week look like? Did that dentist appointment land on the family calendar or mine? (When I’d rather see it than ask, she draws it — the generative-UI trick I wrote about last week.)
And, like most of us, a lot of my calendar begins as email. Boy Scouts logistics for my son, day-camp signups, a conference invite — it never stops. I forward it all to Zora, and she works out where it belongs (family, mine, work, or some mix). Suddenly, it’s just there, without me opening anything. She decides whether it’s even a real commitment, so a newsletter that mentions a date doesn’t get to generate an event; and she can only write to the calendars I pointed her at, never my whole account.
The one I’d actually sell you on began as a specific ask: “before every meeting, build me a briefing — who’s attending, our history, and whatever they’ve been up to lately — and put it where I’ll see it.” I was tired of walking in cold. Now, about a day ahead, she pulls the outside attendees off the invite, looks each one up in my own relationship notes (a private memory I’ll come back to in a minute), then runs fresh web and social searches so I’ve got a real opener when I sit down — “Saw you just shipped that,” or, “How was Tokyo?” She digs through my Gmail, Slack, and Drive for our history and the live thread, finds the agenda wherever it’s hiding — the invite body, a linked doc, an email chain, a PDF read page by page — and writes 500 to 800 words into a “side event” parked right beside the real one.
There are two bugs worth telling you about. Early on, an invite that read SF: Amy | John | Raffi sent the research off on the bare name “John.” It was meant to be John Dickerson, the CEO of Mozilla.ai, who I was about to meet (along with Amy Keating, Mozilla’s Chief Business Officer). Instead it came back, completely sure of itself, with John Lilly, a former CEO of Mozilla Corporation. (The new rule: Figure out who someone is from my own notes first; the web is only for what’s new.) And once, a cheaper model running the overnight job announced “done, six meetings researched” and went quiet, having written nothing. Nine meetings sat empty until I noticed. So now every silent job has to prove it did the work — “show me any briefing still under 500 words” — and raise an error if it can’t. This isn’t only happening on my cheap model: even the big labs’ AI Overviews leave about one in nine claims unsupported by the very pages they cite. Never trust a model that says it’s finished. Make it show you.
#3 Whatever she needs to maintain a private Wikipedia of my whole life
Everyone wishes they kept a CRM for their personal life. Nobody does, because the upkeep is brutal. So I asked Zora to “keep track of the people, companies, and places I deal with, because I never will” — and now she does, quietly, as a side effect of talking to her.
What I have is a Wikipedia-style page for every person, company, and place in my life — not generic bios scraped off the web, but what I actually know: what we discussed, what they’re building, the restaurant they swore by. I mention someone in passing, whether it’s in a chat with Zora or it’s in an email somewhere, and it’s recorded. I forward an Instagram reel about a place and it’s filed, by city, with a “wishlist” tag. Underneath, it’s nothing exotic — a pile of plain markdown files she writes and keeps current. (I pointed Obsidian at the folder so I can read it nicely, but Obsidian does none of the work; it’s just a window onto plain text.) The folder is a git repo that commits itself every few minutes, so there’s always an undo and nothing is ever lost.
Then I ask one plain question — what do we know about this person? that deal? that place? — and she fans out across Slack, email, our old chats, my notes, and the vault, and hands back an answer. I never tell her where to look.
The fun part? It syncs into my actual iPhone Contacts. Open a card in the stock Contacts app and the notes already hold who they are, their title and company, when we last spoke and how often, a couple of recent summaries — all pulled from the vault, kept fresh, no special app. My phone suddenly remembers my people better than I do. And I never typed any of it in. The knowledge just accreted, on its own, in the place I’d actually look for it.
A side effect I didn’t expect: doing this finally cleaned up my address book. Years of duplicates across nearly 7,000 cards, the same person under three names, a heap of empty “ghost” cards from some old sync glitch — she works through all of it. But matching people is genuinely hard, and there’s a long tail she can’t call on her own. So we have a ritual: once a day, in Slack, she sends a short batch of questions — is this the same Sarah as that one? whose number is this? — and folds my answers back in, so the cleanup keeps moving without me ever sitting down to do it. The rule that lets me trust her with the whole book: She only ever adds. She never overwrites a name, a photo, or a note I put there myself.
She also keeps a mirror of the whole thing as a Google Sheet for when I’d rather scan a table than thumb through cards.
I stopped operating. I started asking.
A confession: I didn’t choose these three tasks. I asked Zora what she actually does for me, and the list from her own logs ran much longer, including a weekly research digest off arXiv tuned to my taste; a paper-trading lab running crypto simulations while I sleep, no real money, keeping its own leaderboard; even a rack of Android phones on my desk she taps through to read social media for me. So I picked three. The point of the rest is just this: The web stopped being a place I go, and quietly became a place Zora goes for me.
Here’s the actual skill I picked up this year, and it’s smaller and stranger than “prompt engineering”: I stopped operating software and started asking for outcomes. The move wasn’t clever prompts; it was noticing the things I do over and over — the logging, the calendar wrangling, the who-is-this-again — and simply describing each one until Zora could take it off my plate. Every workflow in this piece began as one sentence. She built the rest and fixed it with me, one mistake at a time. The new literacy isn’t clicking faster or learning the menus. It’s knowing what to ask for.
This is what the melt feels like from the inside. I’m not standing in a room full of forms anymore. I name what I want and the room assembles itself, in my voice, for an audience of one. Hold that picture; it’s about to matter to anyone who builds those rooms for a living.
The harder story isn’t that the web or the app is dying — it’s that all three forces from the top of this piece are coming true at once, and not only for me. Here’s where they braid, and where they stop being about my lunch:
The thing that acts for you has moved up a layer. For 30 years, it was the browser, and a lot of us spent our careers keeping it open and yours. Now it’s the agent — same job, new shape. And that agent is your new front door, increasingly the only one your customer walks through. Your funnel, your storefront, your beautiful app — all of it assumed a human on the other side, reading and tapping. Swap in an agent and the question stops being, “Is my app nice?” and becomes, “Can an agent find me, read me, act on me at all?” Be legible to agents or go dark. Watch how MyFitnessPal answered the call: It shut its API. Zora walked in the side door anyway, on terms MyFitnessPal can’t set and can’t see. Close your front door and the agents come through the window — the only way to be reached on your own terms is to be open on purpose.
The software didn’t get replaced by better software — it melted. Every job Zora took over was an app, and not one got replaced by another app. A purpose-built app was always just a room with a form in it: a database and a handful of actions behind a door you had to walk to. Once an agent builds that room on demand, in your voice, for an audience of one, the reason to download someone else’s version and stand inside it gets very thin. So if you sell software, sit with this: the room was never the asset. You were always selling the database and the actions underneath — the agent just stripped the walls away and showed you what you were really charging for. Stop defending the room. Become the capability the agent reaches for.
And underneath both, the one that decides who wins: whoever holds the agent holds the keys. An agent built on a single closed lab is a rented house. They own the model your business runs on — and can deprecate it out from under you, the way models already get quietly switched off. They own the memory. They hold the keys. The model itself is the boring part now, a commodity current in the wires. The durable value is the harness, the skills your agent piles up by living your work, and whoever you trust to hold them. That’s the case for open, made not as a principle but as a balance sheet: open isn’t the noble choice, it’s the ownable one.
Enough About Me. What Changes Lie Ahead for You?
I have three plain sentences for three types of reader. Build apps for a living? The room you sell is about to be free; decide which capability you are before someone’s agent decides for you. Run a business? The agent is coming through your front door whether you open it or not, and the only version you’ll ever own runs on rails you control: open models, open interfaces, your data on your side of the firewall. Invest? The order of operations is the whole thesis — models commoditized first, apps are next, and the value pools into the harness and the trust around it, which is exactly where the open bets sit.
Because that’s the question under all of it, the one that finally decides everything: whose keys is your agent carrying, and who does it answer to? The thing that walks through every door in your life wearing your name — texts back, fills the cart, schedules your calendar, knows your people — should answer to you, not your landlord. Put the same question to a company and it barely changes: When the agent is your front door, your product, and your keys all at once, is your business standing on ground you own or paying rent to whoever’s property you built it on? The reason my data never crosses my own firewall is the reason your company’s shouldn’t cross a lab’s.
I won’t hand you the business models here — those are the next few pieces. This one plants the flag: the forces are real, I’ve watched all three land on one machine in my closet, and there’s a way through them you can own. The rest of the arc walks it.
The browser was the first thing that went out and acted for me. The agent is the next. The only questions about it that ever mattered — for me, for you, for whatever you’re building — is whether it’s open, and whose it is.
If you’ve got a Hermes agent — or any of the other harnesses — you can build any of this today. Build one, or anything, and show me. I’ll run it, and if it’s good, I’ll write it up here for the people who’d build on it, fund it, or bet a company on it. @raffihack, the comments, or DM.