In short: AI agents do more than automate work. As humans domesticate AI with prompts, evals, and workflows, AI is also domesticating us by taking over the first move: the first outline, the first judgment, the first messy sentence, the first uncomfortable question.
Wheat, Humans, and the Direction of Domestication
In Sapiens: A Brief History of Humankind, Yuval Noah Harari makes a slightly uncomfortable point about the agricultural revolution: perhaps humans did not domesticate wheat so much as wheat domesticated humans. It sounds like a clever reversal at first, but the accounting is fairly plain. Wheat started as a wild grass in the Middle East. Over time it spread across the world, occupied enormous amounts of land, and got humans to clear fields, bend their backs, pull weeds, dig channels, build granaries, and stop wandering. Wheat did well. Human backs, less so.
That story is useful before talking about AI, because it cuts through a lot of vague language about technology changing the world. A tool is not always something you use and then put back on the table. Stay with it long enough and it starts changing your movements, your schedule, and your sense of what feels normal. Wheat changed posture and settlement. The internet changed attention. AI is reaching a little further inward. It is changing how we begin to think about things.
Of course we are domesticating AI. We write prompts, build evals, set rules, and adjust workflows so the model behaves. That part is visible. The other side is less visible: AI is domesticating us into the kind of people it can help more easily. Before a document is open, we ask it for an outline. Before we read the code, we ask it to explain the code. Before an idea has even surfaced, we ask it for angles. Every step makes sense. I do this too. The problem is that after a while, the beginning no longer belongs to you.
This is not a dramatic failure mode. Most of the time it looks like an ordinary afternoon. You sit down, your fingers move before your mind does, and you type a prompt. A few seconds later there is a decent structure on the screen. You relax and start editing. Nothing broke. The work even got faster. You just skipped one more encounter with the feeling of not knowing where to start.
What Changes Is the Order
I find it more useful to think about AI in terms of order than replacement. Replacement sounds loud: a job disappears, a person is pushed out by a machine. The more common change is quieter. The first step changes hands.
Take planning a project. Before, you might draw a few ugly lines on a whiteboard, say something half-formed, and let a colleague push back. Now it is easy to ask the model for three approaches. The result is usually not bad. The headings are neat, the risks are named, and the options look reasonable. But the meeting has already been shaped. People start choosing between A, B, and C. Fewer people go back to the earlier question: was this the right way to frame the problem at all?
Code has the same issue. Old systems always contain things that look stupid from the outside: a strange if, a flag nobody wants to delete, a roundabout compatibility path. AI may tell you this can be simplified. Maybe it can. Maybe it cannot. The trouble is that old code is often not weird because everyone before you was incompetent. Sometimes it is a scar from an outage, a customer, a half-forgotten release. If you only read the summary, you miss the moment of touching the scar.
Writing makes this even easier to see. A first draft is ugly. Sometimes it is ugly enough that the author does not want to look at it. AI is very good at cleaning that stage up. It smooths sentences, adds transitions, and makes the thing feel like an essay. But the first draft was not just trash. It was the first place where you could see what you were actually trying to say. If the model always rounds the language before you meet your own thought, you may meet that thought later and later. Sometimes not at all. You only get a smooth piece.
So I am not saying: do not use AI. That is too easy, and not honest. The thing to watch is whether you are giving away the first step for months or years at a time. Once the first step is gone, you can still edit, filter, and judge. But you are judging inside a shape someone else gave you. Increasingly, that someone is a machine.
This connects to a concern I wrote about in Is AI Making Us Give Up Too Soon?. AI does not only change what we know. It changes how long we are willing to stay with being stuck. Ten minutes of confusion used to be normal. Now, after ten seconds without an answer, it can feel as if something is broken. A lot of ability grows in those ten minutes. Not every time. But enough that losing them matters.
Baumol’s Effect Is Not Sentimentality
If the argument stops here, it becomes the usual advice to “think for yourself.” True enough, but it has the feel of a poster in a meeting room. What makes the issue more interesting is that this is not just cognitive hygiene. It is also about price.
Baumol’s effect is about price. Factories can use machines to raise productivity, so many goods get cheaper. A string quartet does not work like that. Two hundred years ago, four people took roughly half an hour to perform a piece. Today it is still roughly four people and half an hour. Medicine, education, therapy, and care work all contain some version of this. They cannot be compressed like factory output. They do not disappear. In an economy where everything else gets more efficient, they can become relatively more expensive.
AI is bringing this split inside knowledge work. Writing emails, changing tone, organizing notes, translating, generating boilerplate code, drafting standard options: these will keep getting cheaper. They are not useless. They are just easier and easier to produce. In the future, being able to do these things will still matter, but it will not explain why someone is expensive. The machine beside you can do them too, and it does not get tired or annoyed or need tea.
What gets expensive is the part that cannot be made only out of text. A product manager who has personally listened to ten customer calls and then writes a requirement doc may produce something that looks similar to a doc generated from an AI summary. Similar on the surface, at least. But the first person may remember that one user paused before saying “it’s fine,” or that another user did not complain about the interface but simply went quiet for three seconds. Those details do not always fit neatly into a spreadsheet. They still change the decision.
That is the part of Baumol’s effect I care about here. Once AI makes the compressible parts cheap, the uncompressed feel of being there becomes more valuable. Not because “being there” sounds noble, but because it is actually scarce. If you were not there, you do not have it. If you did not talk to that person, you do not have it. If you did not pay some price for the judgment, you do not have that either.
This also connects to another piece I wrote, 5,000 Feeds, 20 Highlights: Your AI Agent Is Killing Your Serendipity. An agent that compresses 5,000 items into 20 highlights seems to be saving time. It is also deciding what will never enter your mind. Many things that change your judgment do not look relevant at first. A stray link, a metaphor from a field you do not follow, an interview you would not normally click. AI is good at optimizing inside known preferences. Human judgment sometimes thickens by getting lost.
So friction should not be deleted as a category. Some friction is drudgery: formatting, moving fields around, cleaning meeting notes. Give that to AI. But some friction is training, some is accident, and some is just the cost of touching reality yourself. In workflow language, all of them look inefficient. Remove them too cleanly and the work becomes smooth. The person becomes thinner.
Don’t Let the Tool Live the Whole Beginning for You
In practice, I think this means keeping a few clumsy habits. Before writing, write a small paragraph yourself. It does not need to be good or even complete. Before making a decision, write down your first reaction, then let AI find counterarguments. When reading something important, read at least part of the original before asking for a summary. Before a meeting, if something feels wrong, write down the feeling even if the reason has not formed yet.
There is no ritual here, and it does not need to be turned into a framework. It is more like leaving yourself a little unprocessed time. Two minutes is fine. Ten minutes is fine. The point is not the duration. The point is whether the problem touched you before it touched the model.
Responsibility works the same way. AI can help with a decision, but it cannot become the person whose name a customer remembers. It cannot explain the judgment three months later. In a few years, many people will be good at using AI. That will become basic competence. Rarer, maybe, will be the person who can still say: this is my call, and if it is wrong, I will explain it. That sounds old-fashioned. In a world where everyone can use the model as a cushion, old-fashioned things may start to stand out again.
Relationships too. Many questions are easier to ask AI now. It is fast, comfortable, and does not make things awkward. But the weight of a relationship often comes from awkwardness: someone misunderstands you and you explain again; you say the wrong thing and repair it two days later; a project breaks and you clean it up together. AI can talk with you for a long time. It has not gone through something with you. Without shared experience, there is no real relationship capital.
Back to Harari’s wheat. Wheat was not smarter than humans. It did not invent a conspiracy. It changed ordinary movements: where people lived, when they woke up, what they ate, what they guarded, why they bent down. Do that long enough and people change.
AI is changing ordinary movements too. Especially the beginning: how we start writing, start reading, start judging, start asking for help. The more natural this becomes, the less it feels like anything is happening. A few years later, you may still be productive, professional, and good with tools. Some things that used to live in you may simply be used less often.
So the conclusion is not to use less AI. That is not realistic, and not very useful. It is closer to this: do not let AI live the whole beginning for you. Let it search, organize, generate, test, and polish. But in a few important places, keep a little rough judgment for yourself. It may be overturned quickly. It may be badly written. Fine. At least the problem passed through you first.