Agency Against Drift

The thermodynamic basis of intention

Leave anything alone and watch what happens. Coffee cools to room temperature. Gases disperse until they fill their container uniformly. Buildings crumble, signals degrade, gradients flatten. None of this requires an explanation in terms of purpose, because none of it is purposive. Under ordinary boundary conditions, systems tend toward the macrostates available in overwhelmingly more microscopic ways. I call the uncontrolled trajectory drift — not a new force or a synonym for entropy, but the baseline evolution a control process must act against.

But things do intervene. A hand reaches out and reheats the coffee. A maintenance crew repoints the brickwork. A cell pumps ions against their gradient; an engineer error-corrects the signal; a mind holds a plan intact against distraction. Locally and temporarily, in scattered pockets of the universe, the slide toward disorder is not merely halted but reversed — and always by the same kind of thing: a system that models its possible futures and spends energy to get the one it prefers.

That kind of system is an agent, and this volume takes its central capacity seriously as a physical process. Energy, entropy, and information already give us rigorous ways to describe parts of what an agent does. Agency — the capacity to model alternatives and make action depend on their evaluation — is usually filed under psychology or treated as an illusion laid over the mechanics. My thesis is that this separation is a mistake. Whatever else agency may be, it is implemented by physical systems, consumes physical resources, and changes physical outcomes. The stronger proposal developed next is that one aspect of it — distributional control — can be quantified.

What Agency Is

Here is the working definition the rest of the framework builds on:

Agency is the physically implemented capacity of an embedded system to model reachable futures, evaluate them, and make its behavior depend on that evaluation so as to change the distribution of outcomes relative to an uncontrolled baseline.

The agent is embedded: it is part of the physical world it acts on, not an observer outside it, and everything it does is done with physical resources from within. It possesses an internal model of reachable futures: agency requires some action-guiding registration of how things could go, though the model may be as spare as a bacterium’s chemical control circuitry. It evaluates: not necessarily in words or conscious feeling, but through an internal organization that makes some modeled states action-guiding and others not. It enacts behavior: models alone move nothing. And the difference is measured relative to a baseline, because an outcome cannot be credited to control until we specify what the same system and environment would have done without the intervention.

This is the book’s basal, control-theoretic sense of agency. Later volumes ask when control also becomes authorship: when a temporally continuous system can own reasons, revise commitments, and preserve the evaluative structure through which its actions remain its own. They then ask when that author is sovereign, rather than instrumentally subordinated to another optimizer. Those are additional layers, not replacements for this definition. A bacterium may instantiate minimal agency without reflective authorship; a reflective author may still lack sovereignty. Causal capacity alone is lower still: a hurricane changes outcomes but models and evaluates none of them.

The Quantum Branching Universe (QBU), developed in its own chapter, supplies one interpretation of that changed distribution: the agent’s action becomes correlated with different Measure-weighted futures. Nothing here requires an agent to edit the universal wavefunction from outside or manufacture branches. The framework’s stronger claim is that causal comparison between an action and an appropriate no-action baseline can quantify how much outcome distribution the agent controlled. That claim will require more than thermodynamics alone; it requires a causal model and, in the QBU version, a defensible account of branching and Measure.

Notice what the definition does not say. It does not say agency is free, and it does not say agency is mystical. Agency is neither. It is a species of thermodynamic work.

The Price of Steering

Physical control requires a physical implementation. Every structured state an agent creates or maintains — the reheated coffee, the repaired wall, the ion gradient, the executed plan — depends on machinery that senses, computes, actuates, or repairs, and that machinery draws on free energy. This does not yet give a universal exchange rate between “amount of intention” and joules. Landauer’s principle fixes a lower bound for logically irreversible information erasure under specified conditions; extending that bound to arbitrary outcome steering requires additional assumptions about the baseline distribution and the physical protocol. The next chapter therefore introduces the kybit first as a proposed unit of distributional control, then states the restricted conditions under which that information measure can also bound work.

Seen this way, intentional action is implemented through an energetic budget, even when the relationship between a high-level choice and its minimum cost is not one-to-one. Three consequences motivate the candidate laws developed next:

• Agency is limited and finite. Every real agent has bounded information, computation, actuation, time, and free energy.
• Agency cannot persist indefinitely without resource throughput. Isolate an agent from usable gradients and its total remaining capacity to sense, model, and intervene eventually runs down.
• Absolute control is physically unavailable. Noise, incomplete models, finite resources, and the cost of correction keep every agent short of a frictionless limit.

These consequences are developed into the three laws of agency, where the robust physical constraints are separated from the framework’s stronger quantitative conjectures. Together they imply that agency is physically bounded, dependent on continuing organization, and therefore an accomplishment.

The Entropy Misconception

Casting agency as a struggle against entropy invites a familiar misreading: the gloomy cosmology in which the Second Law dooms everything and life is a brief candle guttering in a universal wind-down. That picture rests on a confusion about what the Second Law governs.

Howard Bloom makes the case against entropy-pessimism as vividly as anyone. In The Case of the Sexual Cosmos¹ he argues that nature is not cautious or conservative but exuberant — that life thrives precisely by throwing energy at the wall, generating extravagantly wasteful strategies that succeed through redundancy, variability, and flamboyant experimentation: billions of sperm cells competing for a single egg, endless varieties of flowering plants evolving ever-showier lures for pollinators. On this he is right. Apparent wastefulness is often an adaptive strategy; flamboyance is not a violation of nature’s economy but one of its most successful policies. Where Bloom overreaches is in concluding that the Second Law itself should therefore be thrown out. That is rhetorical flourish, not physics — the Second Law is among the most thoroughly verified regularities we have, and it is integral to physics, chemistry, and biology alike.

The resolution is the distinction between isolated and open systems. The Second Law constrains the total entropy of an isolated system; Earth is not one. It receives comparatively low-entropy solar radiation and emits higher-entropy radiation to space, and that throughput makes local order-building possible. When a plant converts available energy into structured organic molecules, its local decrease in entropy violates nothing; the total process still produces entropy. Ecosystems, societies, organisms, minds — every pocket of persistent structure works inside gradients of this kind.

This is the physical setting agency inhabits, not a loophole in the law. Agents are open systems: they exploit free-energy gradients, spend resources on control, and export entropy to their surroundings. Bloom’s flamboyant nature and the austere bookkeeping of the Second Law are not rivals. Creativity can be locally extravagant precisely because the larger process still pays its thermodynamic bill. Agency is that navigation organized around modeled alternatives.

The Stakes

Take the framework’s premise seriously and the traditional picture of choice inverts. Intention is not a ghostly exception to physical law; it is matter modeling alternatives and using physical machinery to make action depend on the model. Drift names the comparison trajectory. Agency names the organized departure. The next two chapters ask how far that departure can be quantified: first the kybit, then the three laws that separate the physical bounds from the framework’s open conjectures.