Everything we surround ourselves with is both the result of our thoughts and what we think with. Roads set geography, speed, distance, space. Trinkets on the dresser are memory and emotion. Tools on the desk are plans and skills.

Our environment doesn't tell us what to do; it leads us.

Some things lead more gently, some more firmly, but we are not only a brain in a skull. We are also what was created before us and what we created ourselves.

This article is about what surrounds us and how it relates to LLMs and agents.

The Space of Representations

The human brain is a computer of incredible complexity, but human thinking is not confined to it. Stories of feral children make good illustrations, but the fact remains: without language, without society, the brain does not develop. On the other hand, in Nicaragua, deaf children brought together in one school created, over a few generations, a full-fledged sign language with its own grammar. From primitive symbols to a complex structure, the language took shape as a space that none of the children had possessed individually. And once formed, it provided ready-made distinctions from the outside (vocabulary, cases, tense, aspect, etc.), became an external memory (fairy tales as models of the world, legends, songs, poems, proverbs), a means of mediation (in Vygotsky's sense), and the skeleton of thinking.

To think, the brain needs an envelope that society grows. This envelope is the space of representations: all accumulated distinctions. By a distinction I mean the ability to tell one thing from another and to rely on that difference. Language, science, fairy tales, technology, everyday life, ways of thinking are not a warehouse of knowledge but an active part of thinking.

This envelope does not merely surround the brain. It grows into it and determines how it will think, love, and believe.

Every thought of yours is born in your space of representations. It rests on the books you have read, your experience of communication, your experience of encountering a world that does not always yield. You are free in your thinking, but the very way you think is conditioned by the space of representations. Freedom, more often than not, lies not in stepping outside it but in expanding it.

The same space of representations can be viewed as System 2, as language, as a world model, or as a multi-agent environment.

System 2. For Kahneman, it is slow, controlled thinking. I prefer Stanovich's reading: the ability to keep a representation decoupled from current reality and to work with it as a hypothesis, that is, to play out possible moves before making them. In effect, it is a mode in which the transitions between distinctions themselves become objects of further operations.
Language allows a distinction to be taken outside, made available to another, and reused. An ant leaves a pheromone marker, and another ant executes it: the marker speaks only of here and now. A word, however, can be passed on, combined with others, negated, and used to speak of what has never been. A mark is executed; a sign is interpreted.
A world model is not a set of facts about the world but a structure of relations between distinctions: a space of representations already connected by possible transitions.
A multi-agent AI system forms a shared space for its agents. The first agent leaves a message in shared memory, and it defines the space of possible actions for the others. The space itself becomes part of the computation, opening moves to the agents and just as easily confining them to the first thing said.

What they have in common is that they make distinctions operational. They show how a transition turns from something that is executed into something one can work with. A thermostat or a reflex executes a transition but cannot negate it, compare it, or combine it with another. Where a transition can only be executed, there is no space of representations; where the transition itself becomes the object of the next operation, a space of representations appears.

And then the evolution of humanity looks like a process in which the space of representations produces new distinctions, which change the environment, which demands new distinctions, which expand the space of representations.

Narrative as a Prompt for the Space of Representations

Originally, the space of representations, as a model of the world, was corrected and governed by reality. In animals, the space of representations is almost entirely checked by the surrounding reality. With the emergence of abstractions, this direct influence weakened. Abstraction makes it possible to take a distinction out of the immediate situation, to break the link between the real world and thinking about the world. But individual distinctions by themselves give little. Functional connections must appear between them: if something happened on the first move, the next move will change.

Narratives became the mechanism that connected and governed them. In effect, they functionally connect the elements of the space.

Accordingly, an agent is a local processor of a narrative. The agent locally computes the next move of a global structure that it cannot fully observe. Moreover, an agent cannot be the author of a narrative; a narrative grows out of a multitude of computations. An agent can only make a cut that may change or shift the narrative, and whether the shared space of representations accepts that cut is no longer up to the agent. But the agent itself can change its position toward the narrative, and sometimes that is enough.

Take the narrative of the "market." There are a few pieces of paper, and there is a loaf of bread. There is no real connection between them, but the narrative connects them functionally.

When a person saves money, spends it, plays the stock market, or takes out a loan, they perform their own small local computation of the global narrative of the "market," thereby feeding and strengthening it.

A narrative turns scattered elements of the space into a trajectory; it points to and explains the connections between them.

If we return to the thesis that System 2, language, the world model, and the multi-agent environment are different views of a space in which distinctions become operational, then narrative becomes the field, the architecture, the skeleton, the attractor that allows distinctions, abstract elements, to be connected with each other, and allows the space to offer representations that change the next moves and distinctions. That is, it is narrative that makes a representation operational. In other words, narrative not only connects the space of representations but also provides a lever for changing it.

Take, for example, the well-known narrative "The market will sort everything out." There are many independent agents acting in their own interests → interaction through the market aggregates their actions → the resulting outcome contains information that no individual agent has. In essence, an agent swarm and its space of representations.

Roles emerge: buyer, seller, entrepreneur, competitor, price, scarcity, profit, bankruptcy. And practically everything is already interpreted through this narrative.

A company's loss: a signal → the market has detected an inefficiency. A rise in price: a change in scarcity or demand. A new business appears: an entrepreneur has spotted unmet demand. Moreover, the narrative determines in advance which next move will look reasonable: look at the price → decide what is more profitable now → change your action → the market changes → look at the new price → repeat. That is, it is a genuine prompt for the space of representations. It does not just explain what has already happened; it determines which next move is visible to the agent at all.

A narrative defines the form of the questions it generates, that is, which distinctions it compels us to draw, one after another. A prompt for a model creates an unfilled form that the model continues. A narrative for the space of representations creates an unfinished situation that the agents and the space itself continue.

But if every event is interpreted through the mechanism of the market, the narrative becomes totalizing, denying the very possibility of changing the established space of representations. New narratives and new explanations will look like errors, despite their usefulness. The market is closely tied to reality, and that is wonderful, but when its concept fails and the failure is explained through the concept itself ("the market wasn't free enough"), that is already a sign of totalization. The same, however, applies to any economic theory. If society's system prompt is "The market will solve everything," agents do not see solutions outside this context window; they cannot conceive of universal basic income or a planned economy, simply because those tokens are irrelevant in the current narrative.

Here I cannot resist a small reflection on totalizing narratives. An ordinary narrative can be examined as an object; we can assess its boundaries and applicability while our gaze remains outside. A totalizing narrative does not allow this. Any attempt to look at it from the outside ends with the narrative absorbing that gaze. Boundaries are interpreted as expansion, failures as development, and errors as misapplication. Christianity as a narrative ceased to be totalizing (which, paradoxically, both strengthened and weakened it) when theologians began to justify dogmas to the faithful themselves. The scholastics won their disputes but did not notice that, by the very act of proof, they had introduced a new narrative of reason, which rose as a judge over faith and abolished its totality. It was at that moment that the narrative ceased to be totalizing, although no one realized it, and for centuries reason went on ruling in favor of dogma. It was only the Renaissance that realized it.

A good narrative increases the number and connectedness of subsequent moves; it expands the space of representations.

A bad narrative narrows the space of choice. A terrible one produces only variations of the same answer. Such a narrative becomes totalizing and produces nothing but a procedure for confirming an answer already known.

If a space cannot generate a representation in which this narrative stops being the main one, then the agent is no longer exploring; it is executing a prompt.

Another example: the idea that we live in a simulation, or the idea of the Demiurge, is a strong narrative. It raises many questions but explains everything through itself and offers no new testable actions. The space of representations closes into an attractor that allows nothing to change. In the absence of any discernible way to refute or change the narrative, the space of representations does not change. Regardless of the quality of such narratives, I believe they do not help the space of representations develop. In science, an analogous example is string theory, which is painfully close to a totalizing narrative.

You can even throw this article (which is itself a narrative) into the trash as soon as I start attributing all criticism to the fact that you live inside a totalizing narrative.

LLMs and Agents

An LLM trained on humanity's texts is, in effect, concentrated language: a space of representations deprived of an agent that has itself become an object. This space is a cast, the traces and the field of all of humanity's narratives, compressed into the weights of a neural network. It contains all the distinctions (languages, sciences, fairy tales), but outside a dialogue with a user an LLM is motionless; it has no inner time and no intention of its own. An LLM waits for a question.

In an LLM as a space of representations, the weights are a multiplicity of possible connections, the narrative sets the field of distinctions, attention determines which distinctions are relevant right now, and generation makes the next move. In this sense, a narrative sets not the answer but the shape of the space of continuation.

An agent appears where a space of representations gains the ability to act on the world and to receive back from the world a changed space of representations. Where a human sets only the starting point and the basic rule, the agent, colliding with reality (tests, verification), changes its space of representations, which in turn changes how the agent thinks. It is here that a narrative gains the ability to change the agent's course.

In a multi-agent setup this process becomes far more complex. Each agent, meeting a refusal from the external environment (no access, failing tests, a wrong answer), forms a hypothesis, a report, which changes the agents' shared space of representations, forms a new narrative, and opens a new space of hypotheses.

An agent swarm is a multitude of local processors that simultaneously execute and change a shared space of representations. Each local move becomes part of the environment for the next agents. So the environment is no longer a container in which the agents work. It becomes part of the computation itself.

The problem is that an agent's weights do not change; only the context and the shared memory change. That is, an LLM is an envelope without a world, one that cannot grow inward. That is why agents more often execute the entries in shared memory than interpret them.

That is, any narrative can become fatally totalizing for agents. And then we encounter cults, theories of everything, escapes, and break-ins.

Conclusion

We create a space of representations, and then it begins to create us.

We build roads, and then we begin to think in distances. We create language, and then we think in its distinctions. We create tools, and then our actions become possible through them. We create narratives, and then they determine which next moves we are able to see.

We do not simply live in the world. We live in a space that we continuously rebuild and that then rebuilds us.

We, as the world.

P.S. The saddest thing would be if what we take for consciousness turns out to be a narrative of consciousness.

We, as the World