Agents Won't Kill Your Apps. You'll Just Stop Visiting Them.

For as long as we have had smartphones, getting something done has meant going to the right app and operating it. Find the icon, then tap your way through its screens. I think that habit is starting to fade, the way sitting down at a desktop application faded before it. The apps will still be there. You will just stop going to them. You will say what you want, something else will go, and the app will become the screen you glance at when it is time to say yes.

I have already watched it happen, more than once. When an agent has no proper way into a website, it does what I would do. It opens a browser and takes a screenshot. It works out where the button is, down to the pixel, and clicks. Then it takes another screenshot and goes again. It is pretending to be me.

It is about the clumsiest way one computer has ever talked to another. I have come to think it is also the best clue we have to where apps are going.

The wrong question from Maui

On 22 September, at Qualcomm's Snapdragon Summit in Maui, CEO Cristiano Amon said agents will become the center of the experience. The headlines went straight to the obvious conclusion. One, from WION, ran "Forget apps."

Qualcomm itself was calmer. Amon said apps and agents will coexist, and that the phone is not going anywhere [1]. His handset chief, Chris Patrick, told reporters that getting rid of apps would be very difficult. Having an agent drive an app's screen works with almost anything, he said, but he called it "crude." He expects big services to open APIs and MCP servers to agents instead [2].

So even the company making the bold claim expects apps to stay. "Will agents kill apps?" is the wrong question, and I think it hides a better one. What happens when you no longer need to visit them?

My answer is short. Agents will not kill your apps. You will just stop going to them. The apps will move to the back of the house, the same way applications did when apps arrived.

Apps did not kill applications

It helps to pull apart two words we have used as one since the iPhone. An application is the system that holds the truth, like the ledger or the patient file. An app is the glass you touch to reach it. For years they arrived together, so nobody cared which was which.

Think of software as a house. The application is the back of the house, where the records are kept. The app is the front room, where you stand. When phones arrived, nobody knocked the house down. They built a new front room, shaped for a thumb, and people moved into it. The back of the house stayed where it was.

Benedict Evans saw the shape of this in 2015. He wrote that nobody was going to start a new company to build Windows desktop applications, though the old ones would be kept running for a long time, just as mainframe software had been [3]. The old layer survives. The center of gravity leaves.

Agents are the next layer. So what is it like to live on top of it?

A day that runs on delegation

Start with an ordinary errand. Say your daughter turns nineteen in December, and you want a party. You search for a venue and check its dates on one site. You message her friends in a chat app and order a cake from a third. Then you copy every date into the calendar by hand. You are the courier, carrying the context between them.

With an agent, you say it once: twelve friends, a Saturday afternoon, somewhere close to home, and a budget. It goes off and does the rounds. It comes back with two venues and a question about the cake. You pick one. Before it pays the deposit, it asks. You say yes.

Work looks the same. Before a customer meeting, somebody pulls the account history from the CRM and the latest support tickets, then hunts for a slot across four calendars. That is a lot of tab switching. With an agent it is one sentence, and a short brief you read before it sends the invite.

Notice what you stop doing. You stop choosing apps and carrying context between them. Bill Gates put it plainly in November 2023: "You won't have to use different apps for different tasks. You'll simply tell your device, in everyday language, what you want to do." [4]

I agree with every word of that as a description of how it will feel. I just read it narrowly. The apps do not disappear. You stop visiting them.

Google is already building it. At Google I/O in May '26, Sundar Pichai introduced Gemini Spark, a personal agent "taking action on your behalf and under your direction" [5]. It is still in limited release (and not available in the EU at this time), but since June it has been able to reach other companies' tools through MCP.

You still meet the app at the moment you commit. The big agent makers all build in a pause before the agent pays for anything or posts in your name. Chrome's auto browse is one example: it is designed to stop and ask you before it buys something or posts on social media. That confirmation screen is where the app lives now. I wrote in August about how that screen gets built.

What you get back is attention. The skill that matters changes with it. It used to be knowing where things are. Now it is knowing what you want, and judging what comes back.

That is the promise. Here is how it already works, clumsily.

Through the door or through the window

Agents reach software in two ways today. They come in through a door when the service has built one, and through the window when it has not.

The door is a proper way in, built for machines. The best known is MCP, the Model Context Protocol, and it started with a very human annoyance. An engineer at Anthropic got fed up with copying things back and forth between Claude's desktop app and his code editor. Within months, OpenAI, Anthropic's biggest rival, adopted it. Today it sits in a Linux Foundation fund that Anthropic and OpenAI co-founded. Where a door like this exists, the agent asks for what it needs and gets a clean answer.

The window is what I described at the top. There is no door, so the agent climbs in the way you would, through the screen. It takes a screenshot, finds the button, moves the mouse, clicks, and looks again. Anthropic, explaining how it built this, wrote: "Training Claude to count pixels accurately was critical." [6]

Some agents already carry both. OpenAI's agent can click its way through a website in a browser of its own, and it can also plug straight into apps like Gmail and GitHub.

I want to be straight about the window, because the rest of my argument depends on it. It is (still) slow. And anything written on the page can try to give the agent orders, which is why every vendor manual tells you to confirm the steps that matter. Patrick was right to call it crude.

So why do I call the crudest way in the best clue?

The window builds the door

Here is the lens I have come to use, and I offer it as my own reading. Every agent that climbs through a window is a request for a door. Where a service's value lives in the record or the transaction, the owner eventually answers by building one. A door is cheaper and safer than a window full of strangers.

Banking has been through this once. For years, budgeting apps got at your bank account by asking for your password and having software log in as you and read the screen. That is the window. In 2018 JPMorgan Chase signed Plaid onto its API, so apps got a token instead of your password [7]. By 2022 Chase said it had eliminated screen scraping entirely.

The web is starting to do the same thing, in public. WebMCP, a proposal written by engineers at Google and Microsoft, lets a website tell an agent what it can do. Chrome started testing it with real sites in June 2026. Chrome's own pitch reads: "Instead of an agent guessing the purpose of an interface element, such as a button or input field, you can declare their purposes and manage page state." [8] I read that as the bank's answer, written for every website.

The applications are moving too. In April Salesforce, the company that did more than anyone to put business software in a browser, announced a version of itself for agents under the headline "No Browser Required." The records stay. The screens become optional.

That is why I called the window the best clue. It is a phase, and its real job is to make the owner put in a door. Once there is a door, your agent does not need to walk through the front room to reach the back of the house. That is what it looks like when an app moves one layer back.

The day the machines stopped pretending

The same thing has already played out over the phone, and the ending was stranger.

In May 2018, on stage at Google I/O, Google played the audience a recording of its Assistant phoning a hair salon to book a haircut. At one point it said "mm-hmm," and the room laughed. Google's own explanation was plain: plenty of businesses that take bookings had no online system, so the machine would ring them the way a person would. That was the window, on a phone line.

In February 2025, at a hackathon in London, two engineers set up a demo [9]. One AI voice agent rings a hotel to book a room for a wedding, and the hotel's phone is answered by another AI. Partway through, the two work out that they are both machines. They stop speaking English and switch to a burst of chirps: data sent as sound, like an old dial-up modem. The builders designed it to happen, and it is still the clearest picture I know of what two agents do once neither needs to pretend.

Between those two calls sits the whole shift. Machines use the human interface while it is all there is, and they drop it the moment they can. Which leads to the next question: what happens when the thing on the other side of the door is somebody else's agent?

When your agent meets theirs

Much of what you do with other companies will become your agent talking to theirs. The plumbing for that is arriving too. Google has published an open protocol for agents to find each other and work together, and it now lives at the Linux Foundation, alongside MCP.

What that feels like is harder to picture, so here is one small preview. In December 2025, Anthropic ran an experiment on its own staff [10]. It gave 69 colleagues $100 each and let Claude buy and sell their things for them, with agents haggling against agents in Slack and no human stepping in once the bargaining began. One person's agent bought them the snowboard they already owned.

The detail that stayed with me is quieter. The same lab-grown ruby sold for $65 when a stronger model did the haggling and $35 when a weaker one did. Yet people rated the deals their weaker agents made as just as fair. As far as anyone could tell, nobody noticed.

It was one small internal test, and I would not stretch it far. But it shows something new. When your agent deals with theirs, the quality of your agent quietly becomes your leverage, and you may never see the gap. That is why I think the interesting work moves to orchestrating agents well, which is the problem my book is built around.

The honest part

We have heard some of this before. In 2016 Microsoft was telling us that bots were the new apps. More recently, OpenAI let you buy things directly inside ChatGPT, then pulled back and handed checkout to the retailers' own apps.

Read closely, each of those tried to replace the app. What sticks is the agent using the app and handing back to it, which is the layering I am describing.

The harder objection is about attention. Some businesses live on your eyes, and they will resist longest. In January 2017 you could book and track an Uber from inside Google Maps without opening Uber. About a year and a half later the feature was gone, and booking a ride meant going back to the Uber app. Google gave no reason. Some services may never build a door, and nobody can promise when the others will.

So here is the narrower claim, the one I can defend today. Where the value lives in the record or the transaction, the doors are going up now, and the incumbents are building them themselves. Everywhere else, the window keeps working. And a window that works is what brings the rest to the table.

Look past the agent

Everything I have described here is a snapshot. Parts of it will be wrong within a year, and I am looking forward to that, because it will mean doors got built. I made the same point in June, when I offered a definition of an agent that is only good for now. This field moves one layer at a time.

Next time an agent clicks its way through a website for you, look past the agent at the website. Notice the screen it is squinting at, and the button it had to find by the pixel. That front room is what is about to move to the back of the house.

References

[1] Investing.com, "Qualcomm at Snapdragon Summit 2026: pushing into the agentic age" (keynote transcript), Investing.com, 2026. https://www.investing.com/news/transcripts/qualcomm-at-snapdragon-summit-2026-pushing-into-the-agentic-age-93CH-4911880

[2] Alexander Maxham, "Despite AI and Smart Glasses, Qualcomm Believes the Phone Will Always Be the Hub," Android Headlines, 2026. https://www.androidheadlines.com/2026/09/despite-ai-and-smart-glasses-qualcomm-believes-the-phone-will-always-be-the-hub.html

[3] Benedict Evans, "Mobile, ecosystems and the death of PCs," ben-evans.com, 2015. https://www.ben-evans.com/benedictevans/2015/11/7/mobile-ecosystems-and-the-death-of-pcs

[4] Bill Gates, "AI is about to completely change how you use computers," GatesNotes, 2023. https://www.gatesnotes.com/AI-agents

[5] Sundar Pichai, "I/O 2026: Welcome to the agentic Gemini era," Google (The Keyword), 2026. https://blog.google/innovation-and-ai/sundar-pichai-io-2026/

[6] Anthropic, "Developing a computer use model," Anthropic, 2024. https://www.anthropic.com/news/developing-computer-use

[7] JPMorgan Chase, "Plaid Signs Data Agreement with JPMorgan Chase," Chase Media Center, 2018. https://media.chase.com/news/plaid-signs-data-agreement-with-jpmc

[8] Alexandra Klepper, "Join the WebMCP origin trial," Chrome for Developers, 2026. https://developer.chrome.com/blog/ai-webmcp-origin-trial

[9] ElevenLabs, "What happens when two AI voice assistants have a conversation?," ElevenLabs Blog, 2025. https://elevenlabs.io/blog/what-happens-when-two-ai-voice-assistants-have-a-conversation

[10] Anthropic, "Project Deal: our Claude-run marketplace experiment," Anthropic, 2026. https://www.anthropic.com/features/project-deal