Agency is goal‑referenced causality. An agent is a system with a world model and a goal state, whose evolution tends to update its world model toward the goal. In this view, “free will” stops meaning acausal magic and starts meaning goal‑referenced causality.
Essay 4 of 5. Previously: prediction is compressed modeling from a limited vantage point, with principled ceilings. Now we add a goal.
We’ve been building toward something, and I can finally say what. Determinism and causality gave us clean vocabulary. Information conservation turned “it was in the cards” into honest bookkeeping. Predictability separated “encoded” from “foreseeable.” One bridge remains before we can talk about free will without contradicting ourselves, and building it is this essay’s job. The bridge is called agency.
The definition (deliberately minimal)
An agent, in the minimal sense this series needs, is a physical system with three features:
Read that again and notice what’s missing: no “chooses,” no “acts,” no “wants.” That’s deliberate. Choosing and acting are the very intuitions we’re trying to explain, so they are not allowed into the definition — that would be circularity, smuggled in through the service entrance. (You might think “tends” is a wiggle word, but it’s actually a term that can be made mathematically precise, which will be a separate deep dive.) An agent, in this bare-bones sense, is defined by the structure of the process: model, goal, goal-referenced updating. Nothing more.
Agency = goal-referenced causality.
This is where counterfactuals become not just passive correlations (”if X then Y”), but the active predictions (”if I do X then Y will happen”).
In praise of the humble thermostat
By this definition, a thermostat is an agent. This might seem like a drastic oversimplification, but look at what the thermostat actually has, because it’s more instructive than it seems.
We casually say a thermostat “maintains the temperature of the room.” But the thermostat has no access to “the room.” It has access to its sensor reading — one number. Its entire accessible reality is that measurement; its perception of reality is not reality itself. (A predicament you may find familiar: recall the “rendered feed” of perception from the previous essay.) Its world model is one number — the reading. Its goal state is another number — the set point. Its updates push the one toward the other. That’s it, and that’s enough.
The moral: a world model does not have to be an inner virtual-reality theater. It can be a single trembling variable.
<aside> 😈
Interlude: The Demon Meets a Thermostat
The demon — its ledger balanced, but its box unsealable and its prediction notebook bloated, unwieldy, and still a work in progress — was wandering the halls of an office building when it met a small beige box on the wall. “What do you know?” asked the demon, out of professional habit.
“65,” said the thermostat.
“I once knew every speck in the universe,” said the demon, “or at least I thought I did. Anyway, today I’m feeling generous, so what is it that you want?”
“70,” said the thermostat.
The demon laughed for the first time since the moon had upset its perfectly ordered box. But while it laughed, the thermostat clicked. The furnace woke, and the room warmed.
The demon stopped laughing. In all its centuries of knowing the position of every particle in its box, it had never once made any of them do anything. “Teach me,” said the demon.
“Click,” said the thermostat, as the furnace shut off — which was actually all it knew how to say.
</aside>
The same goes for a task-bounded AI agent. Its world model might be a small scratchpad — “grocery list requested; peanut allergy; prefers cheap meals” — and its goal state another compact picture: “a completed list satisfying those constraints.” Both models tiny; updating goal-referenced; no self-maintaining organism anywhere in sight.