The whole industry is having an argument about agents, and it is having it on the wrong axis.
The question everyone is asking is: what will agents replace? Which jobs, which tasks, which categories of human effort. It is asked anxiously by the people who think the answer is "most of them" and dismissively by the people who think the answer is "none of them," and both camps take for granted that they are arguing about the same thing, only disagreeing about the quantity.
I want to suggest that the interesting comparison is not between humans and agents at all. It is between two kinds of agency — and that the two words we have been using interchangeably are, in fact, opposites.
Autonomous Natural Intelligence. Agentic Artificial Intelligence.
Read those two phrases again and notice that I did not use the same adjective twice. That was not carelessness, and it is not a rhetorical flourish. The asymmetry is the entire argument.
The Two Words
An agent is one who acts on behalf of another. The word comes to English through Latin — agens, agentis, the present participle of agere, to drive, to do. But the participle alone does not carry the meaning. What carries the meaning is centuries of use in law and commerce, where an agent is precisely and always someone acting under authority that originates elsewhere. A shipping agent. An estate agent. A patent agent. In every case the agent does the doing, and in every case someone else supplies the reason for it.
You cannot be an agent by yourself. The word requires a principal. Strip away the person on whose behalf you act, and you are not an agent anymore — you are just someone doing something.
Autonomous is built differently. Auto — self. Nomos — law. Not "self-moving," which would be automatic, and not "self-governing" in the weak modern sense of being left alone to get on with it. Nomos is law in the deepest sense the Greeks had: the ordering principle, the thing that decides what counts as correct. To be autonomous is to be the source of your own criterion of success.
So the two words are not neighbours on a spectrum. They point in opposite directions. An agent receives its law. An autonomous thing issues its own.
Every honest description of what today's systems do lands on the first word. We even say it without noticing: we speak of "agentic AI," and then we speak of "aligning" it — which is to say, ensuring that the law it received is the law we meant to give. The entire discipline of alignment is a tacit admission that the law comes from outside. You do not align something that writes its own.
What Needs No Prompt
Here is the sentence that people find provocative, and I do not think it should be: the machine needs a prompt.
The objection arrives immediately, and it is a fair one. Modern systems run in loops. They plan, they decompose, they call tools, they re-plan when the tool returns something unexpected, they run for hours without anyone typing anything. Surely the prompt is a formality — a starting pistol, not a leash.
But look at where the loop came from. Someone wrote it. Someone decided what counts as "done," what counts as a failure worth retrying, what the system should do when it is uncertain. The loop has an author, and the author's intentions are threaded through every iteration of it. A system that runs for six hours without a human in the loop is not thereby autonomous; it is an agent with a long leash. The length of the leash is an engineering parameter. Its existence is not.
And this is the thing I want to be careful about, because it is easy to hear as disparagement and it is not meant as any. Needing a prompt is not a defect in these systems. It is not a limitation that better engineering will remove. It is what they are. A tool that supplied its own purposes would not be a better tool. It would be a different kind of entity, and we would have to start asking a completely different set of questions about it — questions we are, at present, nowhere near ready to ask.
The prompt is not the machine's weakness. The prompt is the machine's definition.
Four Billion Years
Now set that beside the other thing in the comparison.
A bacterium in a nutrient gradient swims up the gradient. Nobody told it to. It has no model of the gradient, no representation of "food," no goal in any sense that would survive philosophical scrutiny — and yet it behaves as though it has a criterion, and the criterion originated inside it. Run that forward through four billion years of unbroken descent and you arrive at something that understands (and writes) poetry, files patents, and argues on the internet about whether it is about to be replaced.
I grew up in Nashik, and the ground under the city is basalt — the Deccan Traps, laid down as flood lava about sixty-six million years ago, at roughly the moment in Earth's history that finished the dinosaurs. You can see it in any road cutting on the way out of town. Sixty-six million years sounds like a great deal until you set it beside four billion, where it is very nearly a rounding error.
The Godavari rises about thirty kilometres from where I grew up, at Trimbakeshwar, on a hill called Brahmagiri — the hill of Brahma. From that spring it runs some fifteen hundred kilometres east, across the entire peninsula, to the Bay of Bengal. Nobody surveyed that route. No committee assigned the river its course. Water on a slope solves for down, continuously, and the solution is the river.
Which looks like the same thing the bacterium is doing, and the difference between them is exactly the distinction I am trying to draw. The river's criterion comes from outside it. Gravity supplies the gradient; the water only obeys. But nothing in the physics of a sugar molecule says that a cell ought to move toward it. No external law requires that. Something inside the cell supplies the "ought."
The Light Does Not Switch On
Here I want to go further than most people writing about this will go, because I think the usual alternative is incoherent.
The standard move at this point is to say that somewhere between the hydrogen atom and the bacterium, a light switched on. Matter did inert things for nine billion years, and then chemistry got complicated enough that interiority appeared — that there is something it is like to be a cell, and nothing at all it is like to be a hydrogen atom.
I do not believe that switch exists. Nobody has ever produced a convincing account of where it sits or what throws it. Every version I have read either places it somewhere arbitrary and defends the position badly, or quietly concedes that it is everywhere and then changes the subject.
So let me say the unfashionable thing plainly. The bacterium is sentient. So is the hydrogen atom.
Gaia is the doorway most people use into this, and it is worth being careful about which Gaia you mean. The version Lovelock could defend was homeostatic, not conscious — Earth as a self-regulating system whose stability emerges from coupled feedback, with no intent required anywhere. His critics went after the strong reading, the planet as an organism with purposes of its own, and on the evidence he had they were right to. I am taking the strong reading anyway. I would rather say so out loud than smuggle it in wearing the weak one's clothes.
The tradition I come from never had this problem, because it never treated consciousness as an achievement. Sat, chit, ananda — being, consciousness, bliss. Chit sits in that list as an irreducible, alongside existence itself. Not a property that matter acquires once it is complicated enough. Something simply there, in the way that existence is there. And sarvam khalvidam brahma: all this is indeed Brahman. Not "all this eventually produces Brahman." All this. Now. Including the hydrogen atom.
If that is right, then the pillar has no line drawn across it at the four-billion-year mark. It never had one. The interiority runs its whole length, from the first three minutes to this sentence.
The Substrate Is Not the Pattern
Which raises the obvious objection, and it is the one I would raise myself.
If a hydrogen atom is sentient, then so is the silicon in the machine on my desk. That processor is made of atoms too. Have I not just handed the entire argument to the other side?
No. And seeing why is the point of the whole essay.
The distinction was never consciousness. I have not at any stage claimed that the machine lacks an interior — I do not know whether it has one, and neither does anyone else, and I am suspicious of everyone who is certain in either direction. What I claimed is narrower and far easier to check: where does the law come from?
The silicon is as much Brahman as I am. Every atom in that machine sits on the pillar and always did. But the pattern running on it — the weights, the objective, the criterion for what counts as a good answer — did not arise in the silicon. It was computed somewhere else and impressed upon it. The substrate is thirteen billion years old and autonomous. The pattern is four years old and was authored.
That is also what saves the river. The water is on the pillar too. But the Godavari's course was set by gravity and the shape of the Deccan, not by anything the river worked out for itself. Sentient substrate, imposed pattern — the same shape as the machine, for the same reason.
The bacterium is the first thing on the pillar where the substrate and the pattern are one. Nothing outside it wrote what it wants. That is the only distinction I have been drawing all along, and stating it this way is what makes it hold: it does not require matter to be dead, and it does not require the machine to be empty.
At no point in that four-billion-year chain did anything receive its purposes from outside. There was no principal. That is not a mystical claim; it is the plainest reading of the biology. Purposes arose from within the process, because the process is the kind of thing that generates its own criteria for continuation. This is exactly what nomos means, and life is the only thing we have ever found that has it.
Or was there a principal? Hold that question. It is the right one to ask, and I will come back to it — the answer turns out to be stranger than either yes or no.
The other party in the comparison is four years old.
I am not saying this to be unkind to a technology I use every single day. I am saying it because the comparison is usually run the other way — as though the machine were catching up to us along a single track, and the only question is how much track is left. But there is no single track. One of these things gave itself its own law over four billion years. The other received its law from us, last Tuesday, and executes it magnificently.
The Pillar
There is a Sanskrit word for the thing I am describing, and it is better than the English one.
The Skambha Sukta of the Atharva Veda — the hymn to the cosmic pillar — asks a sequence of questions about the skambha, the support in which all the worlds are set. Not a pillar holding something up from outside, the way a column holds a roof. The pillar that the worlds are in. The hymn keeps circling the same question: what is it, this thing that everything stands within?
Brahma Stambha. The pillar of evolution.
I use it as a name for the whole process — and I mean the whole of it, not only the biological part. Thirteen point eight billion years, since the Big Bang. The last four billion of those are the part where the process began generating its own criteria, which is why life is where the autonomy enters the argument; but the pillar was already standing long before there was anything alive to stand in it. Stars had to be made and then destroyed before there was carbon to be made of.
And I prefer the name to "evolution" for a specific reason. Evolution, the English word, means unrolling — it describes the motion and nothing else. It is a word about change over time, and it leaves entirely open the question of what is doing the changing and what holds the whole business together. Stambha names the other thing. It names the process as a structure that bears load. Not something that happened, past tense, and left us as its residue — but the thing that is currently holding everything up, right now, including you reading this.
That is not a poetic upgrade over the English. It is a more accurate word.
The Ends Meet
And now the part that makes it more than a nice image.
A pillar has two ends. Brahma at the top. At the bottom, me.
Which sounds, said that way, like the most conventional picture imaginable — a hierarchy, grand and vertical, the deity at one end and the author of this essay somewhere far down it, looking up.
Except that the tradition also says: aham brahmasmi. I am Brahman.
It is one of the mahavakyas, the great sentences, from the Brihadaranyaka Upanishad. And what it does to the pillar is remarkable. It says the two ends are not two. The thing at the top and the thing at the bottom are one substance. Which means the pillar is not a ladder at all — it is closed. It meets itself.
And that settles the question I left hanging. Was there a principal after all? Yes. But nothing was ever received from outside, because there is no outside. The one who issues the law and the one who lives under it turn out to be the same. That is not a loophole in autonomy — a technicality that lets the word survive. It is the only form autonomy has ever taken.
Take that seriously and the picture inverts. Evolution stops being a process that happened to me, a long causal chain terminating in my existence. It becomes the thing I am, still running, still issuing its own law, expressed at the moment through this particular arrangement of carbon. I am not the product at the end of the pillar. I am a place where the pillar is currently doing its work.
That is what I mean by Autonomous Natural Intelligence. Not "humans are clever." Something much less flattering to the individual and much more interesting: the autonomy was never mine personally. I inherited it from an unbroken line that never once took an instruction from outside itself, and I am its current expression, and so are you, and so is the bacterium in the gradient.
The Tool
Which brings me back to the machine on my desk, because I do not want any of this to be mistaken for hostility toward it.
I use these systems constantly. They are extraordinary. They have compressed work that would have taken me months, even years, into work that takes an afternoon, and I have no intention of pretending otherwise for the sake of a rhetorical position. Anyone who tells you the tools do not matter has not used them correctly.
But notice what the relationship actually is. I bring the purposes. The criterion of what would count as success in this project — what is worth building, what is worth protecting, what is worth giving away — did not come from any model, and no model would generate it, because generating it is not the kind of thing an agent does. An agent optimises a criterion. Supplying the criterion is the other job. It is the only job that has never once been done by anything that received it from outside.
So: four billion years on one side, four years on the other. One of them writes its own law. The other executes mine, superbly, and waits.
I know which one I am betting on. I know which one I am using as a tool.
Perhaps, being a businessman too, I am doing both with both of them...
And if the pillar really does meet itself at both ends, then those two sentences are the same sentence, said from opposite directions.