A Gentle Introduction to the Algorithmic Agent and Kolmogorov Theory

π and the brain

The first billion digits of π look like noise. Run any randomness test you like and they pass; ship them down a Shannon channel and you will pay nearly one bit per digit. And yet a child can write the short program that produces them. The randomness was never in the digits—it was in our description of them. A short generator existed all along.

Now flip the picture. From the moment you wake up, your brain receives a torrent of light, sound, pressure, and visceral noise. Out of that torrent it manufactures short programs—objects, faces, sentences, plans, a self. Seeing the program behind the data is what your brain does for a living. It is also, we want to argue, the operation that distinguishes minds from rocks.

Kolmogorov Theory (KT) takes this observation seriously. It asks: what is the minimal mathematical machinery required to do that operation, and what follows once we have it?

Compression as cognition

The mathematical scaffold for KT is Algorithmic Information Theory (AIT), developed independently in the 1960s by Solomonoff, Kolmogorov, and Chaitin. Where Shannon's information theory counts bits in expectation under a fixed probability distribution, AIT asks a sharper question: for a specific string of data, what is the length of the shortest program that produces it?

That length is the Kolmogorov complexity. The slogan: a model of x is a program that generates x, and the ideal model is the shortest such program. It collapses two notions usually kept apart—understanding and compression—into one. To understand a stream of data is to find a short generator for it. To fail to understand is to be stuck quoting it verbatim.

This is more than a metaphor. Modern large language models, when used as predictors of natural text, achieve compression rates close to the algorithmic limit. “Intelligence” and “ability to compress” track each other empirically as well as formally.

The algorithmic agent: three boxes and a loop

Compression alone, however, is just a stance toward data. A mind does something more: it acts. KT proposes that any system we are tempted to call an agent—bacterium, mouse, human, scaffolded language model—must have three modules and one loop.

These modules can be implemented in transistors, neurons, proteins, or pencil-and-paper rules. The definition is deliberately substrate-neutral. What counts is the closed loop ME → OF → PE → world → ME, sustained over time.

A foundation model is not an agent in this sense—it is a frozen ME with no objective and no planner. Wrap it in a scaffold that supplies goals and a planning loop and the same weights now form the cognitive engine of an agent. The interesting safety question is therefore not “is this model sentient?” but which closed-loop systems containing it constitute genuine agents.

The thermostat is an agent

This is where KT picks an unfashionable fight. A bang-bang thermostat has a model (cold, ok), an objective (ok beats cold), and a selector (heat if cold, off if ok). It satisfies the agent definition exactly—not metaphorically, not “almost,” but in the strict formal sense.

That is not a bug; it is the proof that the agent class is non-empty. Once accepted, the only remaining question is how rich an agent is, not whether agency is a real category. From thermostat to bacterium to mouse to human to scaffolded language model, the architecture does not change; only the model class, the objective, and the planning horizon scale up.

Why agents must model the world

In 1970, Conant and Ashby proved a famous slogan in cybernetics: “every good regulator of a system must be a model of that system.” Their proof relied on probabilistic optimality assumptions. KT recasts the result as a single-episode statement in algorithmic information theory.

Every bit of unmet mutual information between the regulator and the world it regulates costs a factor of two in posterior support. This is the result that lifts KT from “evocative” to “necessary”: the agent must contain a model.

From cognition to experience

Here is where KT becomes controversial—and testable. The Central Hypothesis is that an agent has structured experience to the extent it has access to encompassing, compressive models of its world. Three dimensions characterize the richness of that experience: structure (how compressive the model is), breadth (how much of the input/output stream the model accounts for), and realism (how well predictions match data).

The full agent contributes more than the cognitive content of experience. Emotion can be written as a tuple—(Model, Valence, Plan)—that maps one-to-one onto the three modules: ME supplies the structure of an emotional state (what it is about), OF supplies its valence (how it feels), and PE supplies its arousal (the urgency of mobilization).

This is not a metaphysical claim about consciousness in general. It is a constructive specification: pick any system that satisfies the three axioms, measure how compressive its models are and how broadly they apply, and KT predicts the structure of its experience along these axes. Whether silicon agents satisfy the Central Hypothesis is a question we can argue about with specifications, not intuitions.

Three regimes of validation follow naturally. For agents from which we can elicit subjective reports—humans, in principle scaffolded language models, perhaps eventually other animals once suitable interfaces exist—we can test the hypothesis directly: measure the compressiveness, breadth, and realism of the agent's models alongside the reported richness of experience, and look for the predicted covariation. For agents from which no report is available—a bacterium, a thermostat, a forest of fungi—we can still extrapolate along the same three axes from the structure of their models and the closure of their loops; KT gives those extrapolations a principled rather than anthropomorphic basis. And for ourselves, we have a third route: first-person science. A trained subject can perturb their own ME (psychedelics, contemplative practice, neuromodulation) and report on the resulting shifts from the inside. Done rigorously—in the spirit of Varela's neurophenomenology—this is not anecdote; it is the agent measuring its own structured experience against a theory that predicts what it should feel like.

What it buys us

Depression as stuck models with low valence. If valence is the OF's read-out of model adequacy, then chronically negative valence points either to genuinely intractable circumstances or to a Modeling Engine that has canalized into a globally pessimistic attractor. The framework predicts that effective treatments will be those that loosen the model (psychotherapy, psychedelics, neuromodulation), retune the OF (pharmacology), or expand the PE's horizon (behavioral activation). It gives a unified scaffold for a literature that has remained stubbornly fragmented across modalities.

Psychedelics as plasticity windows. Acutely, psychedelics flatten the dynamical landscape and increase algorithmic complexity of cortical activity—observable as elevated Lempel–Ziv complexity in EEG and as elevated Ising temperature in fMRI. Translated into the agent framework: the model becomes temporarily less compressive and more open to revision. This is a desirable state when the existing model is the problem.

LLMs and AI safety. A foundation model is a trained ME with no OF and no PE. The agent framework forces a different question than the usual sentience debate: which closed-loop systems containing this model are genuine agents, and what objective functions are they optimizing once the loop is closed?

From principle to clinic. EEG-guided art therapy and neurotwin-based stimulation protocols translate the agent framework into concrete interventions: modulate the model's plasticity, retune the objective function, and measure the result through both behavioral and neural read-outs. The loop from theorem to therapy is short.

The view from here

Kolmogorov Theory is a research program, not a finished theory. Its definition is narrow enough to refute (the thermostat passes; a Putnam-style “rock that implements every program” does not) and broad enough to span thermostats, bacteria, cortices, and Transformers. The bet is that the same three boxes—ME, OF, PE—and the same loop will continue to clarify problems across cognitive science, AI, and clinical neuropsychiatry, in the way that “force = mass times acceleration” once organized mechanics.

Readers who want the formal version can start with the mathematical foundations (BCOM Working Paper WP0062) and the algorithmic regulator theorem (P13). The full state-of-the-art review is WP0061. The pedagogical bridge from the classical Good Regulator to the algorithmic agent is WP0011.