The Agentic Compressor

A modeling system should not just learn from data — it should plan how to learn, treating model-building itself as a space of strategic choices.

The core intuition is simple but underappreciated. When a physicist builds a model, she doesn't just absorb data and run gradient descent. She asks: which hypothesis is worth testing? Which part of my current model is probably wrong? Should I reparameterize, unify two subtheories, or focus attention on the anomalous data points? These are choices — and they're made strategically, guided by an implicit sense of which move will most improve the model's explanatory power. This paper argues that any serious agent architecture needs to formalize exactly this process.

Within BCOM's Kolmogorov Theory (KT) framework, an agent is decomposed into a Modeling Engine (ME, which builds compressed world models), an Objective Function (OF, which assigns a scalar "valence" to states), and a Planning Engine (PE, which selects actions). Previous work specified what the ME must produce — compressed representations satisfying mutual algorithmic information constraints — but left the process of model construction as a black box. This paper fills that gap by proposing the Agentic Compressor: a ME equipped with its own internal action space (attend, simulate, compare hypotheses, reparameterize, revise submodels, invoke symbolic reasoning, etc.) and an internal planning loop whose objective is compression gain — the reduction in description length achieved by a given model-building move. The agent now runs two coupled planning loops simultaneously: one acting on the world, one acting on its own model.

Transformer attention turns out to be an existence proof that this idea already works in practice. During training, attention dynamically routes gradient signal so that different parameters are updated differently depending on the current example — that's implicit, selective model revision. It's one internal action (attend + revise) implemented differentiably. The Agentic Compressor generalizes this to a full repertoire of model-building operations under an explicit planner, rather than a single implicit mechanism.

The paper's most striking section concerns what happens when this architecture goes wrong. The agent has one Objective Function: telehomeostasis — persist. Normally, compression gain serves this goal, because better models enable better planning. But when no external action can relieve persistent low valence, the same internal planning machinery can be redirected to modify the world model not for accuracy but for affective relief — making the world feel better rather than making the model truer. The paper argues this single mechanism accounts for religious belief, denial, dissociation, and delusion: all instances of the planning engine sacrificing compression quality to keep valence above a viability threshold. Whether this tradeoff is adaptive or pathological depends on whether the resulting model degradation creates a vicious cycle (worse model → worse planning → lower valence → more distortion) or a one-time rescue that restores function.