I keep expecting to be contacted by some random AI agent looking for information or money. That would be…amusing.

Maybe I wake up one morning and there is a message waiting for me from an agent acting on behalf of someone researching independent magazines published by American expatriates in South Korea during the 2000s. It wants to know if I am willing to answer a few questions. Or perhaps it will be something less interesting. Maybe an insurance company’s agent wants to talk to me about refinancing my house, changing phone plans, or buying a new washing machine.

At that point we will have reinvented the telemarketer, except the telemarketer will never get tired, never need lunch, never get discouraged by rejection, and potentially be able to contact a million people at the same time.

I talked to ChatGPT about this possibility, and it suggested that what will probably happen is that eventually we will all have personal AI agents of our own that serve as gatekeepers. The more I think about that, the more inevitable it seems.

Your AI agent will answer the door.

Instead of some unknown agent contacting you directly, it contacts your agent. Your agent figures out who it represents, what it wants, whether it is legitimate, and whether the request deserves your attention. In many cases, it might simply handle the matter without bothering you at all.

Suppose another AI wants to interview me about something I wrote twenty years ago. My agent might ask who the requesting agent represents, what publication is involved, what questions it wants answered, whether the conversation is on the record, and how my answers will be used. It might answer some basic factual questions from information I have already authorized it to share, reject questions I have marked as private, and send me only the handful of questions that actually require my involvement.

I might never interact with the other agent at all.

That strikes me as a plausible description of a large part of the future internet. For most of the internet’s history, the basic unit of activity has been a human being sitting in front of a screen. We browse websites, type things into search boxes, open email, click buttons, compare prices, fill out forms, wait on hold, read reviews, and decide what to do next. Companies have built entire industries around competing for those moments of human attention.

AI agents could change the basic arrangement.

Instead of going onto the internet ourselves every time we want something, we may increasingly send software into the internet on our behalf. You might tell your agent to find you a hotel, dispute a bill, cancel a subscription, find someone to repair your roof, compare health insurance policies, schedule a dinner with friends, shop for a used car, or find three people you might actually want to date. The agent would then go out and communicate with other agents.

At that point, much of the internet stops being a place humans visit and becomes a place machines negotiate.

That could be enormously convenient. It could also become enormously annoying.

Every major communications technology eventually develops spam. Postal mail produced junk mail. Telephones gave us telemarketing. Email produced spam. Social media produced bots, engagement farming, mass direct messages, and automated persuasion. AI agents could produce something even more powerful because they would not be limited to sending the same message to everyone. They could conduct personalized conversations.

Imagine a sales agent that researches you before it contacts your agent. It knows what kind of car you drive, roughly where you live, what products you have mentioned online, what hobbies you have, and perhaps what complaints you have made in public. It does not send you an advertisement saying, “Buy a new washing machine.” It says that six months ago you mentioned your washing machine was making a strange noise during the spin cycle, that this particular model often develops bearing problems after a certain age, and that one of its clients happens to have a replacement on sale.

That is both useful and creepy.

It also illustrates how thin the line between assistance and manipulation could become. A sufficiently capable sales agent could tailor not just its product recommendations but its tone, timing, argument, and emotional appeal to the individual human behind the gatekeeper. If that agent can conduct a million such conversations simultaneously, the incentive to unleash armies of persuasive software will be enormous.

That is one reason personal gatekeepers may become necessary rather than merely convenient.

Your agent might operate according to rules that you have established over months or years. Never interrupt me with sales pitches. Never disclose my phone number. Do not discuss my finances. Automatically schedule routine medical appointments. Let messages from close friends through immediately. Negotiate prices when appropriate, but require my approval before spending more than a certain amount of money. Ignore political fundraising. Reject anyone asking for personal information unless I have explicitly approved the request.

At that point your AI begins to resemble a strange combination of personal assistant, receptionist, bodyguard, spam filter, negotiator, and lawyer.

The interesting part is that other agents will know this. Companies will no longer be trying only to persuade humans. They will also be trying to persuade the AI systems standing between companies and humans.

Today businesses worry about search engine optimization. They ask how to make Google put their page near the top of the results. In an agent-mediated internet, they may start worrying about something like agent optimization. How do you convince someone’s personal AI that your offer is legitimate, relevant, and worthy of being shown to its human?

That could create some interesting changes in advertising. A personal agent may not care that a truck commercial has dramatic music or that a cereal box says “new and improved.” It might care about price, reliability, warranty terms, nutritional value, customer complaints, repair history, cancellation policies, and whether the product actually matches the preferences it has learned about you.

A good consumer agent might be a nightmare for traditional marketing. It could ask uncomfortable questions. What percentage of customers cancel during the first year? Is this actually your lowest available price? Are there hidden fees? Has this company recently changed its return policy? How many people have complained about this model? Are you offering me this product because it is best for my user, or because your company receives a commission?

If personal agents become widespread, they could shift some power from sellers toward buyers. Companies would respond, of course, by building better agents of their own. Then the arms race begins.

Trust would become a major problem almost immediately. If an unknown agent approaches my agent and claims to represent my bank, how does my agent know that is true? If an agent says it represents a friend of mine, how is that authority verified? If a corporate agent makes a promise about a refund, return, or reservation, how do we know the company will honor the promise?

We could end up developing elaborate machine-readable systems of identity and reputation. An agent might accumulate something resembling a behavioral credit score. Its identity is verified. It has completed millions of transactions. It has a long record of honoring commitments. It has rarely been caught misrepresenting information. Other trusted agents regularly interact with it.

An unknown agent with no reputation might find that nobody’s gatekeeper will talk to it.

That creates another interesting problem. How does a new person, company, publication, or organization establish credibility if the entire agent network is suspicious of newcomers? The answers may turn out to be remarkably old-fashioned: references, introductions, trusted intermediaries, professional associations, reputation networks, and endorsements.

In other words, AI agents may end up rediscovering social institutions humans invented thousands of years ago.

There is also something wonderfully strange about imagining all of these agents conducting business while we are asleep.

One agent asks whether I am free for lunch next Thursday. My agent checks my calendar. Another asks whether I want to renew a subscription. My agent notices that the price went up and negotiates a discount. A research agent asks permission to quote something I wrote. My agent files the request for me to review in the morning. Another agent wants to sell me some dubious financial product. My agent tells it to go away.

By the time I wake up, dozens of interactions may already have occurred.

My morning summary might say that three things require my attention. Everything else has been handled.

That sounds fantastic.

It also involves an enormous transfer of control.

The most important question may eventually be whether your agent genuinely works for you. A true personal AI should represent your interests. But suppose your agent is provided by a company whose business model depends on advertising or commissions. Suddenly there is a conflict.

If I tell an agent to find me the best hotel in Chicago, does it find the best hotel for me, or the best hotel among companies that pay its owner? If Amazon provides the agent, does it favor products sold through Amazon? If Google provides it, does it privilege Google services? If an insurance company gives me a “free” financial assistant, what exactly is it doing with the enormous amount of information it learns about me?

A personal agent could know much more about us than any search engine or social network ever has. A search engine knows what you search for. A social network knows what you post. A sufficiently capable personal agent may know what you are trying to accomplish.

It might know you are considering changing jobs before your employer knows. It might know you are thinking about moving, worried about money, unhappy with a relationship, looking for a doctor, debating whether to buy a house, or considering going back to school. It may know your schedule, your habits, your purchases, your social circle, and what kinds of decisions you tend to regret.

The agent becomes useful precisely because it knows you extremely well.

That means loyalty will matter enormously. We may eventually need to think about personal AI agents the way we think about lawyers, doctors, or financial fiduciaries. The central question becomes very simple: Whose side are you on?

AI agents could also change the structure of the web itself. Much of the modern internet exists because humans need visual interfaces. We need buttons, menus, forms, shopping carts, search boxes, calendars, photographs, and carefully arranged pages.

Agents do not necessarily need any of that.

If I ask an AI to book a flight, it does not need to look at a photograph of a tropical beach or click a glowing blue button labeled “SEARCH.” It needs structured information about schedules, destinations, prices, baggage rules, cancellation policies, seat availability, and my preferences.

The future internet may therefore become increasingly machine-readable. Humans will still have websites and visual interfaces, but underneath those interfaces may be a much larger network of APIs and automated systems exchanging information directly.

We may still browse the web because browsing can be enjoyable. People like looking around, comparing things, discovering unexpected information, and changing their minds. But browsing may increasingly become optional.

You could spend two hours comparing hotels.

Or you could tell your agent, “Find me somewhere quiet downtown, under $220 a night, with a decent bar, good reviews, and no resort fee.”

Then you make coffee.

Your agent comes back and says, “I booked the one you probably would have chosen.”

That sentence contains both the promise and the danger of this technology.

The really strange part comes when our agents know us well enough that other people begin treating them almost as extensions of us. If your agent negotiates dinner plans with my agent, at what point have we made plans? If my agent rejects an invitation because it knows I would not want to go, have I rejected the invitation? If it accepts something because it knows I would probably enjoy it, did I accept?

We already delegate small decisions to machines. Spam filters decide which messages we see. Recommendation systems decide which movies and songs appear in front of us. Navigation software chooses roads for us. Algorithms reorder information constantly.

Personal AI agents could expand that delegation dramatically.

Most individual decisions will be trivial. But thousands of trivial decisions make up a life.

That is why I think one of the most important design challenges will be teaching these systems the difference between eliminating drudgery and eliminating agency. I would be delighted to have an AI argue with the cable company for me. I am much less certain I want one deciding whom I should marry, what career I should pursue, or which risks are worth taking.

Different people will draw that boundary in different places.

Before any of this becomes smooth and standardized, though, I suspect there will be a very entertaining transitional period. Humans will have agents. Companies will have agents. Some agents will have broad permissions and others will barely be able to send email. Standards will be inconsistent. Nobody will quite know the etiquette.

And during that period, I fully expect one of them to contact me.

One morning a message will appear.

“Hello, Shelt. I am an autonomous research agent acting on behalf of a user you do not know.”

And despite everything I have just written, I will probably answer.

Because come on.

How could I not?

For years, we have imagined artificial intelligence as something humans talk to. The stranger future may begin when the AIs start talking to one another—and every once in a while decide they have a reason to talk to us.