CALLIOPE · BCOM

A glider in Conway's Game of Life persists. It is not built into the update rule, the lattice carries no preferred frame, and almost every initial condition extinguishes any candidate structure within tens of steps. Yet the glider is identifiable across thousands of generations and reproducible across substrates. The same is true, mutatis mutandis, of a vortex in a turbulent flow, a soliton on a fibre-optic line, a crystal in a melt, a mitochondrion in a cell. The recurrence of persistent patterns in complex dynamical systems is the empirical fact that demands explanation. We argue that the explanation is structural rather than incidental: any admissible persistent pattern in an algorithmic soup, in the precise sense of WP0168 (observer-relative projection of the soup state $X_t$ into a projected description $P^ _t$ whose normalized mutual algorithmic information across time is bounded below by a threshold), is by the Algorithmic Regulator Theorem doing the work of a regulator and therefore admits a refactor into a modeling engine, an objective function, and a planning/action engine. In this precise sense, every persistent pattern is an algorithmic agent. WP0168 introduced this move as a transition from pattern to agent requiring a further ``actionable for the pattern itself'' criterion; we show that criterion is automatic. The proton, the diamond, the glider, the thermostat, E.~coli, and the human brain differ not in whether they are agents but in measurable properties of the agentive refactor: invariance breadth, model dimensionality, update depth, counterfactual horizon, policy richness, objective hierarchy, and self-modification capacity.

Every pattern that genuinely persists through time is, in a precise structural sense, already an agent.

The paper starts from a simple observation: the universe is mostly noise. Run Conway's Game of Life from a random start and almost everything dies within dozens of steps. Yet a glider — a five-cell shape — translates diagonally across the board for millions of generations without the rules ever mentioning it. The same phenomenon appears everywhere: vortices in turbulence, solitons in fiber optics, bacteria dividing for billions of years, cities outlasting every person in them. These patterns persist not because the substrate was designed to support them, but because something structural is going on. This paper argues that structure is agency.

The core move is to invoke the Algorithmic Regulator Theorem (developed in a companion paper, WP0168). The theorem says: if a subsystem reduces the description-length of a world's observable outputs — i.e., if it acts as a regulator — then it must share non-trivial mutual information with that world. Persistence, defined formally as high normalized mutual algorithmic information between a pattern's description at time t and at time t+τ, is exactly what a regulator does. So any pattern that genuinely persists (under a projection that is non-trivial, intelligible, useful, and stable over time) is automatically doing regulatory work. And any regulator can be "refactored" into three functional roles: a Modeling Engine (ME) that tracks the relevant state, an Objective Function (OF) that evaluates continuation of the pattern, and a Planning/Action Engine (PE) that selects behavior. That triple is the definition of an algorithmic agent. The implication is one-way and automatic: persistence forces agency. No extra "actionable for the pattern itself" criterion is needed, contrary to what WP0168 claimed.

This doesn't collapse all distinctions. A proton, a diamond, a glider, a thermostat, E. coli, and a human brain are all agents in this structural sense — but they differ enormously along seven measurable axes: how broadly the pattern survives perturbation (invariance breadth), how rich the internal model is, how deeply the system updates its model from experience, how far ahead it plans, how many distinct actions it can take, whether its objectives are hierarchically structured, and whether it can rewrite its own refactor. These are quantitative differences, not categorical ones. The paper is explicit that calling Schrödinger evolution a "policy" or carbon lattice physics a "planning engine" is a deliberate extension of ordinary usage — the payoff is that the same conceptual architecture applies at every level, and the differences between levels become measurable rather than mysterious.

The paper also addresses why this isn't vacuous. A classic worry (Putnam-style) is that you can always find some mapping that makes any system look like it's computing anything. The response is that admissibility criteria block this: a Putnam-style mapping requires a projection that is not intelligible (its complexity grows with the trajectory length), not useful (it doesn't support prediction), and not persistent (it doesn't survive alternative initial conditions). All three criteria independently rule out the trivial case. The paper also notes an important asymmetry: persistent pattern implies agent, but agent does not imply persistent pattern. An engineered system with an externally imposed reward function is an agent structurally, but its persistence isn't guaranteed by its agency — which the paper identifies as the formal core of the AI alignment problem.