Two people on my team gave the same agent the same brief last week. The results were not remotely the same.
It was an ordinary brief, the kind we hand to a coding agent a dozen times a day: take this rough internal tool, make it something a non-technical person could actually use, here are the constraints. Same model. Same system prompt. Same repository, the same hour, the same documentation open in the next tab. One came back with a thing that worked and felt like a chore — correct, joyless, the software equivalent of a form you fill out at the dentist. The other came back with something a stranger could pick up and understand in ten seconds, with the sharp edges quietly filed off in three places I hadn't even thought to mention. The difference wasn't the machine. The machine was identical. The difference was who had been sitting in the chair.
I have started thinking of the model as a session musician. Not as a metaphor for marketing copy — as the most accurate description I have. A great session player can play anything you put in front of them. Country, funk, a film score, a jingle for regional car insurance. They are fluent, tireless, and astonishingly capable, and they do not care what you ask them to play. Hand them a sloppy chart and they will give you a clean, competent take on a sloppy chart. They will not save you from your own bad song. The record still sounds like whoever was in the producer's chair, because the producer is the one who knew what the song was supposed to be.
That is the part the current panic keeps missing. When the band can play anything, the band stops being the variable. The variable becomes the person who decided what to play.
Same instrument, different room
For most of the history of making things, the bottleneck was capability. Could you write the code, render the image, draft the contract, build the model. The scarce thing was the doing, so the doing was where the value lived and where the differences showed. Two people with the same idea produced different work mostly because one of them could execute it and the other couldn't, or executed it worse.
In 2025 that floor has dropped out from under a whole category of work. The capability is now ambient. It is the same capability for you, for me, for the competitor down the street, available for the price of a subscription, improving on a schedule none of us controls. If access to a powerful model were the edge, there would be no edge, because everyone has the same access. We would all be making the same record.
We are obviously not making the same record. Which means the edge moved. It moved off the instrument and onto the person holding the chart.
I want to be precise about what that person is actually contributing, because "prompting" makes it sound like a trick with magic words, and it isn't. The two briefs in my opening were not different because one person knew a secret phrasing. They were different because one person had a sharper picture, in their own head, of what good looked like. They knew the stranger who would use the tool. They had a flinch about the joyless version that the other person simply didn't have yet. The specification was better because the taste behind it was better. The prompt was just the place that taste became visible.
The prompt is a mirror
I wrote an essay a couple of years ago arguing that taste — the trained, mostly-unconscious sense of which competent thing is actually good — was the scarce thing the generators couldn't supply, and couldn't be delegated. I stand by it, and I think the arrival of capable agents has done something I didn't fully see coming. It hasn't just left taste as the scarce input. It has made taste legible. It writes your judgment down in a form anyone can read.
When you ask a model to do something, the request is a confession. It reveals what you think the problem is, how clearly you can see the desired end state, what you noticed and what you didn't, which trade-offs you even knew were trade-offs. A vague request reveals a vague mind. A request that anticipates the edge case reveals someone who has met the edge case before and remembers it. You cannot fake this with vocabulary, because the model will faithfully execute exactly the level of thought you brought, no more. It is a mirror that talks back. It will hand you mediocrity with the same fluency it hands you the right answer, because, like the session player, it cannot tell which one you actually wanted. Only you can.
There's a quiet comedy in watching this land. People arrived at these tools expecting the machine to supply the thinking, and the machine cheerfully supplied the typing instead, and handed the thinking right back. The work got faster. It did not get easier to be good at.
Authorship survives
The fear underneath all of this is that the work stops being anyone's — that when the same model serves everyone, output converges into one beige average and authorship dissolves. I understand the fear. I think it has it backwards.
When production was expensive, the labor of making hid the author. You couldn't always tell whose judgment was driving, because so much sweat went into the doing that the deciding was buried inside it. Strip the labor away and the deciding is all that's left, naked and unmistakable. The agent removes the part of the work that was never the signature — the rendering, the boilerplate, the fourth revision of the same loop — and leaves the part that always was. The choices. The standard. The flinch.
So I no longer worry that the model erases the author. I think it does the opposite, the way a recording reveals a producer's ear once the engineering is handled. Give a hundred people the same impossibly capable session musician and you do not get one record a hundred times. You get a hundred records, and the good ones belong, recognizably, to the people who knew what they were chasing. My own company is built on putting these agents directly in the hands of people who were never the makers — the ones who used to wait in line for a technical department to have time for them. Watching them work has only made me more sure of this. The tool is the same in every pair of hands. What comes out is not, and the difference is the person, every time.
The uncomfortable corollary is the one worth keeping. If the machine plays whatever you can specify, then there is nowhere left to hide a thin idea or a half-formed standard. You used to be able to blame the bottleneck. Not enough hours, not enough budget, the engineer was busy. Those excuses are evaporating. What you make now is a clean read on what you knew and how clearly you wanted it.
The instrument got infinite. The taste of the person who asked is the whole song.