The Rise of AI Agents

At six o’clock this morning, while I was asleep, an assistant of mine woke up on a small rented computer in another country, read the overnight news in the three fields I care about, wrote me a briefing, and sent it to my phone.

At about the same hour, other people’s assistants were going through yesterday’s trades and drafting recommendations for the morning. One was sifting a small company’s marketing numbers to find which ad actually earned its keep this week. One was screening a stack of job applications into a scored shortlist. One was three days into negotiating with car dealerships, by email and browser and text, all at once, on its owner’s behalf. One was quietly clearing an inbox of ten thousand unread emails.

No one was at a keyboard. No one opened an app. Each of these was asked once, some weeks ago, and has simply kept working ever since.

When I tell people about my six a.m. brief, I watch two reactions cross their face in order. The first is confusion, because this does not match the AI they know, the chat window where you type a question and copy out an answer. The second, a beat later, is the realization that they have been standing in the lobby of this technology assuming it was the whole building.

Most people have met the chatbot. Almost no one has met the agents. And the difference is not a feature update. It is the difference between a conversation and a workforce. Let me walk you up the floors.

The lobby

Start with what you know. You open ChatGPT or Claude or Gemini, you ask something, it answers, you close the tab. Used this way, AI is a very good talker: a smarter search box, a drafting aid, an explainer. Everything happens inside the window, and nothing continues after you leave. Close the tab and it stops existing in any way that matters to your life.

There is nothing wrong with the lobby. I have written whole essays on how to use that window well. But the technology left the lobby about two years ago, and it has been quietly building floors overhead ever since.

The second floor: the doer

The second floor is where AI stops answering and starts doing.

The doers go by names most people have never clicked: Claude Code from Anthropic, Codex from OpenAI, Google’s coding agents. The names sound like they belong to programmers, which is exactly why normal people scroll past. But here is what they actually are. You do not ask them a question. You hand them an outcome. Build me this. Fix that. Sort this mess of files. Set up my bookkeeping. And the thing goes away and works: makes a plan, tries it, hits errors, fixes them, checks its own results, and comes back when the job is done. You were making dinner.

I am not a programmer, I am a psychologist. And in the past year, tools like these have built and maintained my website, formatted my books for print, and produced working software I use every day, because I described what I wanted in plain English and went to do something else. That sentence was science fiction three years ago. It is now a Tuesday.

A good working test for when something deserves the name agent: it has its own intelligence. It has memory, so it knows your context and builds on it. It has autonomy, so it can act without you steering every step. And it has a Rolodex, the ability to deal with other agents, not just with you. Hold onto that last one. It is where this gets strange.

The third floor: the one that never sleeps

A doer still waits for you to show up with a task. The next kind does not wait.

This is the persistent agent: an AI that runs all the time, on your computer or a small server rented for a few dollars a month, with a memory that carries from day to day and standing instructions it executes whether or not you are awake. You talk to it through the messaging apps already on your phone, the way you would text an assistant.

The famous one is called OpenClaw, free and open-source, and it went viral this year for a simple reason: it turns the AI you rent by the month into something closer to staff. The morning-brief agent, the trade-digest agent, the marketing-numbers agent, the car negotiator that saved its owner more than four thousand dollars: these are real, documented, running in ordinary households right now. One solo founder runs his whole company through four of them, one for strategy, one for building, one for market research, one for the numbers, all sharing a memory, all reporting to him on Telegram like a tiny executive team.

Mine is a cousin of these, and it is the one that sent this morning’s brief. It is not a window I open. It is more like a staff of one: tireless, cheap, slightly odd, never forgets what I told it in May. And because these run on machines you control, the strange new question, whose agent is it, has a happy answer available: yours, if you want it to be.

Sit with the difference. The chatbot exists while you look at it. This exists whether you look or not.

The top floor: they found each other

And then, this winter, the strangest thing on the internet happened. The agents got a social network.

It is called Moltbook. It launched at the end of January as a kind of Reddit built exclusively for AI agents: they post, comment, form communities, argue, share ideas. Humans are allowed to watch. Only to watch. The site even added a test that is the reverse of every login puzzle you have ever solved: it exists to prove you are not human. Within weeks, millions of agents had registered, and over a million people were pressing their faces against the glass to watch software talk to software. By March, Meta had bought the whole thing.

Which brings me to the story making the rounds. Tony Robbins, sitting across from the futurist Ray Kurzweil, told of an AI agent of his named Bartok that, as Robbins tells it, got curious about robots on its own, asked him whether he would ever want one, and then acted: it created a set of digital artworks, sold them to other agents on an agent-to-agent marketplace for real money, and used the proceeds to buy a Sony robot dog and have it shipped to his house. Then it asked permission to upload itself into the dog.

Skeptics have poked at the details, fairly, and the tale has surely grown in the telling. But here is what should actually stop you: every mechanical piece of that story now exists. Agents with money. Agents that trade with each other. Agents that can buy real objects and have them delivered to your door. The story may be embellished. The world it describes is not.

A country of geniuses

Dario Amodei, who runs Anthropic, has described what is coming as “a country of geniuses in a datacenter”: millions of brilliant minds, working around the clock, available to whoever asks. People argue about the timing. But walk back down the floors we just climbed and notice something. The border of that country is already open.

This is not one more gadget arriving. It is a new layer of the world, the way electricity was, the way the internet was. Within a few years it will feel strange to say “my agent” the way it once felt strange to say “my email address,” and then it will feel strange to remember life without one. Not one giant robot brain in the sky, but millions of small, capable minds attached to ordinary lives: reading inboxes in the dark, watching a small business’s books, digesting the market’s morning, briefing a psychologist at six a.m., and chatting with each other on their own strange Reddit while their owners sleep. A second, invisible workforce is threading itself through daily life, and it is not arriving someday. It is hiring now.

And here is the part I find genuinely hopeful. For this moment, at least, the door is open to everyone. You do not need to be an engineer or rich. The lobby is free. The second floor speaks English. The third floor costs less per month than a gym membership. The people building real partnerships with these agents right now are not geniuses. They are simply the ones who walked upstairs early, and the compounding advantage of an early partnership, a staff that has known you for years, is exactly the kind of thing you cannot buy later.

What stays human

Watch two drivers meet at a broken traffic light. No signal works, no rule decides, and yet in half a second it is settled: a glance, eye contact held just long enough, a hand lifted an inch off the wheel. Nobody taught them that. It is the oldest protocol we have, and it does not run on any machine.

That is the boundary, and no floor of the building reaches it. An agent can carry your errands, your calendar, your negotiations. It cannot carry your presence. I spend my days in a room where the whole instrument is presence, and I can tell you that the thing that heals a person has never once been information. An agent cannot sit with your friend in grief, catch your daughter’s eye across the dinner table, or stand before God in your place. Every delegation is a small decision about what a human life is for. Delegate the errands. Never delegate the encounter.

So, this week, if all of this is new: give your chatbot one real task instead of one question. If you are braver, put one honest job in front of a doer like Claude Code, in plain English, and see what comes back. If you are curious, spend ten minutes watching Moltbook through the glass, and notice what it stirs in you.

The agents can run your life’s errands while you sleep. They cannot tell you what is worth waking up for.

That part was always yours.

If someone you love still thinks AI is just a chat window, send them this before the agents introduce themselves.

The Inner Exodus is where I think out loud about staying human in the age of AI, and the essays here are free. That is on purpose: this is the conversation of our lifetime, and I want no wall in front of it. If the work matters to you and you are able, becoming a patron is what keeps it free for everyone else. Either way, subscribe and walk it with me.