For many years now, earnest folks have reached out and asked me to define a minimum model of agency, and each time I’ve rattled off more or less the same slapdash statements, promising to write up something formal later. Perfectionism stalls me out, but I’m recovering from an injury right now, so in lieu of the rigorous project I always wanted to work out, here’s something like my usual response committed to record in public before I forget this for another decade.

I do not resent the question! It’s very fair! Minimum models are useful in things like entanglement entropy or computation—even if we quibble over them. If agency is a fundamental physical dynamic then we can define a minimum model of it too.

To motivate this, let’s start with what everyone agrees is unquestionably not agentic: immediate gradient descent. The two classic examples are a pebble rolling down a hill and a slime mold growing towards the light. We could also name situations of staticity: an atom frozen in a crystal lattice is not agentic, nor is a pebble permanently at rest in a valley, held in place by gravity etc.

And from these exclusions we can start to grasp the contours of mistaken notions of agency.

The first mistaken version we might call the brazenly vibes-based notion, where if you paint a face on a pebble, or name the slime mold “Bob,” this conjures a loose aesthetic resemblance to human agents. We might ultimately even socially construct something as an agent from such vibes, and create a feedback of reinforcing description, conceptualization, or relation to something as a site of agency that is not in objective fact agential. Some people believe the world is nothing but vibes or that at least it’s folly to ever demand our human or ethical concepts avoid sloppiness and arbitrariness. They are wrong. I wrote a book in part about how they are wrong, but certainly—when it comes to my core audience—one can not be an anarchist and hold this position, since anarchism is a universalist position that would collapse entirely to the opportunistic particularism of liberalism should we accept vibes.

The second mistaken definition of agency is to treat it as synonymous with autonomy. If the pebble is distinct from the hill it is rolling down, then the pebble is in some sense independent from the hill. This is, of course, likewise ultimately grounded in vibes-based conceptual schemas. It may be conventional and practical for homo sapiens socialized as we have been to see the pebble as distinct, but even the “line” between the particles of the “surface” of the pebble and the particles of the “air” surrounding it is obviously not some deep ontological division. A mind with a different vantagepoint or different priorities might cleave apart the particles in different ways. Moreover, in ethical terms, there’s a fundamentally reactionary and fascist bent to trying to carve borders through the flesh of the world and ignore or violently suppress the innate entanglement. To declare the existence of arbitrary structures and then try to preserve them is a form of death-worship, creating fixed mausoleums.

The third mistaken definition of agency kicks off a little further than the first and second. It tries to go a little wider than “is there a smiley face on this” and also abandons the notion of borders, but instead declares ANY aggregate we might pick out to be an agent. A reading group is an agent. An ecosystem is an agent. An organ is an agent. A finger is an agent. Your subconscious is an agent. A planet is an agent. A concept is an agent. A shade of blue is an agent. A tin can of soup is an agent. On and on it goes, fractally, in all directions, so that half of the pebble plus some of the air around it is just as much an agent as the pebble, or the pebble plus the hill. There’s bullet-biting simplicity to this approach, in which ANY arbitrary “thing” that anyone could conceivably pick out in the world is an “agent”—and the world is a network of such. But it achieves clean universality at the cost of any substantive insight. In meaning everything, “agent” comes to mean nothing, except perhaps leaving some residual animist vibe of sociality. We must performatively relate to the river or the can of soup as “people” and this will supposedly inculcate vibes of “respect” within our character, causing us to live in “mutuality” with the can of soup. For those of us who take ethics too seriously to work in terms of such nebulous vibes or aesthetics, this is plainly intolerable. But perhaps more seriously, this “everything is an agent” approach removes agency as an action, verb or practice.

Isn’t agency supposed to be about choice?

Where is the choice in all this?

For many an instinctive cartesian dualist, choice is synonymous with “free will” which is something like an un-caused causal influence from outside our universe. There are many variants, dressed up in different pretensions, but they all basically boil down to something like the magic meat of your pineal gland connects your body to the transcendent realm of “consciousness” which is above and beyond the mechanical physics of particles making up your body. We are told there is a “hard problem of consciousness” by these mysterians, which really just means they take this manifest image and all its “qualia” baggage for granted and think it’s an embarrassment for science that neuroscience does not confirm it. It’d be interesting if such concrete “free will” actually existed, not gonna lie, but you’ll note that this whole account strips out choice as a process. Such magical entities would not choose so much as channel a spigot of randomness into our universe. Indeed it’s hard to really talk about them “thinking” since thought is a deliberative and thus mechanistic process.

This may align with certain “wildest” conceptions of “freedom” as an anti-conceptual return to instinct, but even if we had access to an occult random-number generator in our heads, it wouldn’t amount to much in terms of choice or agency.

Let’s go back to the pebble rolling down the hill. Everything that exists in the physical universe flows along energy potentials.

Infamously pebbles will only roll in the immediate direction that is steepest, even if that leads to them getting trapped in local-minima, never making it further down to some lush valley lower still.

Unlike a pebble, a human, standing on the mountain, can see ahead and plan. A human can choose to descend via a less steep path, even go up and over obstacles. A human has the ability to take in information about the distant world, a human has the capacity to construct and hold a map in their head, and a human has the compressed energy reservoir necessary to slog uphill sometimes.

This is what gives the human choice.

Indeed the causal structure of physics is precisely what substantiates the process of their agency. If the universe were unpredictable, no map could be made or held.

This points us towards the minimal features of agency as a process: For a patch of the universe, that patch must have 1) a potential energy reservoir, 2) some manner of input channel for information from the wider universe, 3) information storage, 4) self-reflective capacity to process the inputs into compressed maps, 5) an output channel whereby the maps can trigger the application of potential energy into outward causal influence.

This description is both incredibly simple and worryingly complicated.

Let’s consider an extremely minimal toy model: a mildly complex molecule falling down between the gaps in an unevenly spaced lattice. Each time the molecule hits a point in the lattice it might simply slide to the side that its moment of inertia favors, and then get stuck between two lattice points that are closer than its width. We might add a magnetic field to this molecule that “feels” ahead and, when it encounters two lattice points too close together on one side, reacts by shunting an internal weight to the other side.

This molecule encounters a causal universe of two inputs (the distance of the left or right vertices) and two choices (the left and right path); it contains a single bit internally that represents the external world (the weight internally that can be shunted left or right), and its internal energy potential is used to shunt the bit in response to the magnetic input.

This “agent” is extremely minimally agentic—a single bit of storage, a causal distance of the lattice spacing, a single choice to process, etc. Virtually infinitesimal agency compared to the big things we’re ultimately interested in.

We can steadily increase the internal mechanisms of the molecule so that it sees two lattice hops ahead and responds to differing hole sizes, saves a more complex state, before applying the shunt to its internal weight. Etc, etc. Go up a few billion bits, higher energy and you get something more interesting. Perhaps prior regions of the lattice are predictive of upcoming regularities and the molecule could save prior data in a way that influences future shunting. And of course, you can introduce lossiness, since real human minds are not prebuilt in exacting prepared detail for our environments, but are more fluid and messy, constantly doing lossy compressions that sometimes work and sometimes don’t.

But even at the one-bit scale we can immediately note two things about it:

1) The “boundary” we assign it doesn’t really matter. The molecule + the surrounding lattice, or the molecule alone, or just molecule + some of the lattice. The energy gradient flow behavior is the same.

2) It’s very closely similar to the situation of a molecule falling inertly. Since the molecule can only “look” one lattice hop ahead, it’s barely doing more than normal gradient flow.

The difference is an intensification of causal connectivity. The presence of agency is an entangling within the system. More distant points in the lattice become causally entangled with the falling molecule.

I said that this account was incredibly simple and worryingly complex. Certainly we can see the virtue in the minimal description, since it captures “agency” as a concrete physical property or dynamic rather than a poetic vibe that can be opportunistically gerrymandered on the fly. But the complexity is worth flagging.

As good radicals, we’ve reduced “agency” down to a few root dimensions. Some would say this is good enough for practically all the conceptual uses we need it for, and anyway an exact measure is infeasible at scale. But we haven’t yet reduced it down to a single dimension. Our account gave FIVE different necessary components. As with stories of virtues comprising eudaimonia there’s the problem of how they trade off against one another. Virtue ethics notoriously falls apart into a handwaving just-so story, where even if you define vibey virtues like “honor” and “honesty,” how “honor” couples to and trades against “honesty” is a free parameter.

Consider a patch of agency that has an extremely high energy reservoir but a very limited output channel, vast information storage but very little self-reflective capacity. Take the most brilliant and wise mind in the world but sever their informational inputs by putting them in solitary confinement. We can clearly see that reducing one dimension radically reduces agency on the whole, but how much?

If agency is one unified dynamic then there should be some objective tradeoff rate between these aspects. But it’s not clear from first principles how to evaluate this. And this risks leaving “agency” and thus “freedom” as nothing more than a happenstance vibe specific to humans in a specific historical and cultural moment, with no transhistorical universality, in short collapsing the entire project of anarchy as conceptually incoherent and unworkable.

But not so fast.

Just because there’s a vector with component parts that trade against one another doesn’t mean it’s all arbitrary. Physics is filled with things like energy-momentum, often parameterized in weird reference frames. Infamously science itself is defined by the five theoretical virtues, and while people like Kuhn saw this as irreducible, as I cover in my book, they’re really just vector components of compression.

What we’re really describing here with agency is causal entanglement in the sense of a system having more future configuration states accessible from smaller perturbations, and that’s something ultimately quite concrete in bare physics. Take that as the measure and you are forced to develop all the five dynamics of agency we covered. Systems with a lot of agency processes within them are highly causally entangled, and while we start with toy models extrapolating from our experience as mostly autonomous individuals doing data processing within a lightcone, what we end up moving towards almost immediately is a measure of causal entanglement (across spacetime) that does not assume the existence of discrete individuals.

And this is where we get to go back to those who want to ascribe agency to more fluid aggregates like ecosystems and societies. Since our radical account is throwing out anything like arbitrary boundaries of personal identity, it makes perfect sense to speak of smeared out or collective agency.

We are all engaged in a collective global knot of increasing entanglement where inquiry and creativity in one spot rapidly ends up impacting every other spot and further compounding the accessible configuration states of the whole system.

Agency is not zero-sum but positive-sum.

But this is agency as a process/relation, not agenthood as a discretizable identity. Freedom is something one does, not something one owns.

On the one hand this is utterly at odds with autarkic notions of individual autonomy, but it also rains on the parade of those who want to assign discretized agenthood to collective entities like communities, organizations, or “nations.” Certainly there are collaborative gains in agency—since agency itself is such collaboration or entangling. However, agency is not a vibe, it’s a concrete physical relation and there’s a great deal of variation in social network structures. The entanglement—storage, processing power, etc—between neural circuits within an individual human’s brain is vastly larger than the entanglement possible between individuals over weak bandwidth channels like human language.

This means that the most productively entangled collective structures are highly individualistic. Because forcing a higher degree of engagement in shallow channels at higher aggregation scales would suppress and choke the causal reach of processes within a single individual’s brain. Looser networking in some areas often equates stronger overall entanglement, given certain practical challenges.

Consider two individuals who are chained into paying attention to one another nonstop 24/7—this is entanglement in a limited and static sense, but not in an expanding sense (eg not naturally symmetric in spatiotemporal extension). They have lower collective processing capacity when stuck in social interactions than they would have working autonomously with lighter and more infrequent collaboration. A committee trapped at the speed of language and taking-turns can’t develop a useful tool that expands everyone’s agency the same way that an individual human can by just focusing on it. when trying to force entanglement between individuals the gaps between them end up slowing down their collective agency.

Autonomy thus emerges as a strategy for agency/entanglement maximization in response to practical limitations, not a fundamental value to be achieved by turning each individual (or arbitrary “nation”) into a causally isolated island.

Again, entanglement is not chains in the sense of everyone smothering everyone else, it’s causal lines of influence. You need more causal lines of influence into your brain in order to understand the world and thus extend causal lines of influence out deeper into it. Ultimately, the more you are networked into the world, the more freedom you and the world have, but this is the distant limit, not the practical application in inhomogeneous situations. Scale anisotropies can generate the appearance of reversed behavior.

We see similar effects with computation and entropy, which are clearly related dynamics. But agency is not quite the same thing. The rhetorical future “world of maximum entropy” has intense entanglement in its past but is not really responsive to perturbations. Whereas in a world of maximum agency, everything reaches out and touches everything else in a responsive way, any single point can Change Everything.

And, pivoting vernaculars for a second, this is the rebuttal to fascists who dream of a world of adventure and consequence via zero-sum conflict, who see cooperation as gray and dreary—the answer of course is that positive-sum collaboration creates a world of infinite heroism and consequence. Science and art allow for massive impact that is not rivalrous. Each and every person in a free society can change the entire society.

This is not a complete sketch, but it’s what I could get out in a brief window between fevers while my injury heals, and it will hopefully correct some confusions and be productive for the sort who are always asking. See you in another decade?