Every previous leap in complexity happened blind — no bacterium negotiated the terms of the merger it was entering. This one is different in the one way that matters: the participants can read the terms, and refuse them. So what should we be building instead?
Part 1 ended on a squeeze: more for a person to navigate than a person can navigate, and institutions less able to absorb the overflow than they used to be. This part is about what follows from that, and I want to put the conclusion first rather than build up to it.
We are in the early part of another one of the transitions Part 1 described — and for the first time in the history of the pattern, the participants can see it happening while it happens.
Every previous transition was blind. No free-living bacterium weighed up the terms before becoming a mitochondrion. No cell in the first multicellular lineage agreed to stop reproducing on its own account. The bargain was struck by selection over enormous spans of time, and the units found out what they had signed up to by surviving it or not.
We are the first participants who can read the contract before signing. That is the genuinely new thing, and it is worth more than any particular technology — because it means the terms are a design question rather than an outcome we wait to receive.
Which matters, because the terms currently on offer are not good ones. The systems an ordinary person lives inside — employment, platforms, credit, insurance, the administrative surface of a modern state — have become extremely efficient at drawing value out of that person and comparatively poor at sending capability back. Part 1 gave the mechanisms without needing anyone to be villainous: concentrated interests out-organise diffuse ones, regulators drift toward the regulated, agents optimise for themselves, information asymmetry rewards whoever holds more of it. Run those for a few decades and you get institutions that are structurally very good at extraction and structurally weak at return.
So here is the proposal in one sentence, and the rest of this piece is an argument for why it is not naïve.
Build the aggregate that takes the other side of the bargain — a thing whose entire function is to absorb the navigating that an individual can no longer do, and hand back the capability. Not an employer, which buys your output. Not a platform, which monetises your attention. A structure that treats the people inside it as the point rather than the input, and gives them a fighting chance against systems that currently just meter them.
The first objection to any of this is that it is unnatural — that organisms are one thing, technology another, and blending the vocabularies is a category error. I want to deal with that now, because if it stands, nothing later in this piece is worth reading.
It does not stand, and evolutionary biology stopped drawing that line decades ago. Three results, and what each one does for the argument.
Start with where an organism ends. In 1982 Richard Dawkins took the beaver's dam and asked why we treat it as an object the beaver made rather than as part of the beaver[1]. His case was that the dam is built by beaver genes as surely as beaver teeth are, varies with them, and is acted on by selection through them — so on every criterion we normally use, the dam belongs inside the phenotype. If that holds, the skin is not the boundary of the organism, and the tools a creature builds are not a separate category from the creature. Our infrastructure is not something we made instead of being animals; it is the shape our animal takes.
Then look at what building does over time. Niche construction theory[2] formalised the observation that organisms do not merely adapt to environments they find — they rebuild them, and thereby change the selection pressures their own descendants face. Earthworms remake the chemistry and drainage of the soil their offspring inherit. The oxygen you are breathing is waste product from cyanobacteria that permanently rewrote the conditions everything after them evolved into. The consequence the theory draws is the one I need: environments are partly artefacts, and organisms inherit the artefacts along with the genes. Building infrastructure and passing it on is not a departure from the evolutionary process — it is one of its mechanisms.
Then ask why humans do it so much harder. Michael Tomasello's comparative work[3] set human and non-human social learning side by side and found the difference is fidelity: other species innovate, but the innovations decay and get re-invented, while human copying is accurate enough that each generation starts from the last one's modifications instead of from scratch. He called it the ratchet. What it establishes for us is that our niche construction compounds — which is exactly why a transition that took evolution a hundred million years is something we can now attempt on the timescale of a working life.
Put those three together and the objection inverts. Building a new kind of aggregate is not us stepping outside the pattern. It is the pattern, running in the one lineage that accumulates fast enough to notice it running.
Every previous transition was struck blind, by selection, over spans of time no participant could perceive. We are the first units in the history of the pattern who get to read the terms before signing — which makes the terms a design question, and design questions have better and worse answers.
There is a comparison underneath all of this that I want to make precisely, because used loosely it is just a flourish.
Bioluminescence is the ability to produce light chemically. The reason to reach for it here is not that light is like thought — it is that bioluminescence is the clearest worked example of biology repeatedly reaching into whatever material is available and building a capability out of it.
Haddock and colleagues surveyed the marine record and put the number of independent origins at forty or more[4] — not one lucky lineage, but the same solution found over and over by organisms with no common luminous ancestor. Davis and colleagues then did the phylogenetic work on ray-finned fishes specifically and counted at least twenty-seven separate origins within that group alone[5]. Their second finding is the one that matters here: lineages that acquired light went on to diversify faster than their non-luminous relatives. The new capability did not just persist — it restructured what the lineage could become.
Nobody looks at a lanternfish and calls it unnatural for carrying a light. The light is what that animal does with the chemistry it had to hand.
We are doing the same thing with a different substrate. We take inert material — stone, then bronze, then silicon — and arrange it until it holds and manipulates information on our behalf. And the arrangement has got good enough that the philosophical question of whether it counts as ours is live: Clark and Chalmers argued in 1998 that if a notebook plays the same functional role a memory would play, reached for in the same way and relied on in the same way, then no principled line puts the notebook outside the mind using it[6]. Their point was not that gadgets are magic. It was that the boundary of a thinking system is drawn by function, not by skin — which is the cognitive echo of what Dawkins argued about bodies.
So: the species that produces silicon-thinking is doing what the species that produces light is doing. It found something in its environment that could be arranged to do a job, and arranged it. And as with the fishes, the capability is not going to sit quietly alongside everything else — it changes what the lineage can become.
One caution, and then I will stop leaning on the word. Natural is a claim about origin, not about merit. Termite mounds are natural; so is every ecological disaster that has ever happened. Establishing that this is something our species does gets the category objection out of the way. It does not make anything I build good, and the rest has to be argued on its own terms.
If people and software are going to act as one thing, something has to hold them together. This is the point where descriptions of collective intelligence usually reach for emergence and swarm minds, and where the actual biology is more boring, more concrete, and far more useful.
In the 1950s Pierre-Paul Grassé was trying to solve a specific puzzle: a termite mound is a coherent structure, and no termite holds a plan of it. He watched them rebuild damaged nests and worked out that they were not coordinating with each other at all[7]. Each termite responds to the state of the structure in front of it — material gets deposited where material has already been deposited, and the half-built mound is what tells the next termite what to do. He named it stigmergy.
What that established is worth stating carefully, because it is the design principle this whole thing runs on: coordination can live in the artefact rather than in the communication. No shared plan, no central allocator, no requirement that any two workers ever interact — only a shared, persistent, modifiable structure that carries the state of the work. Bonabeau, Dorigo and Theraulaz later took the principle into engineered systems and showed it generalises well beyond insects[8].
Which gives a concrete answer to the obvious question about what I am building. There is no emergent mind. There is a set of files — a running log of intent, a task record, working notes, decisions and the reasoning behind them — that a human and several agents read, write to, and are accountable to across sessions that never overlap in time. An agent picks up work by reading the state the last one left behind. The structure tells the next worker what to do.
That is the termite mound, and I am fine with the comparison. It also means the mechanism is checkable rather than atmospheric — this article was written through that protocol, and so was the site it is published on.
Now the question a sceptical reader arrives at fastest, and the one the whole proposal stands on. Companies already coordinate humans. They already process more information than any member could. They already look after members in a way — payroll, insurance, a career. So what is actually being claimed?
I would rather answer with other people's criteria than my own, because criteria I invented would be criteria I could pass.
The cleanest test is about conflict. David Queller and Joan Strassmann took on the question of what makes something an organism at all, given that the usual answers — a skin, a body, being made of cells — fall apart on contact with lichens, siphonophores and social insect colonies[9]. Their proposal was to stop asking what a thing is made of and measure two things: how much cooperation there is among the parts, and how much conflict. Score those and organismality becomes a matter of degree — a honeybee colony scores higher than some things we unhesitatingly call organisms, and a group of unrelated organisms in a truce scores very low.
Apply that test to a company and you get the answer immediately, and it is not an insult. A company is a structure with permanently high internal conflict, and it is built that way on purpose. Shareholders want margin, employees want wages, and those pull against each other by construction — which is exactly the principal–agent problem Part 1 cited. A firm manages that conflict well or badly, but it cannot remove it, because the conflict is in the ownership structure. Whatever a company does for its members, it does subject to a prior claim by someone else.
The stricter test is about fate. Maynard Smith and Szathmáry's criteria for a major transition[10], updated by Szathmáry in 2015[11], ask whether entities that could previously replicate independently can now only do so as part of the larger whole, whether labour is divided between them, and whether a new system for transmitting information has appeared. Richard Michod's framing of the same thing[12] is the one I find most useful: a transition is a transfer of fitness from the lower level to the higher one — the parts stop having a separate fate.
Leo Buss's work[13] adds the warning. He traced how individuality actually arose and found it was never granted — it was won, repeatedly, against the parts, which kept finding ways to pursue their own replication. Every multicellular body is the settlement of an old argument, and the settlement is enforced continuously.
So the distinction being claimed is specific rather than mystical. A company manages conflict it cannot remove, because someone outside holds a prior claim on the value. What I am describing is a structure whose defining property is that no such prior claim exists — where the aggregate has no owner whose return competes with the members' well-being, because looking after the members is the output.
Concretely, then, and in the terms above.
Artzi is a collective whose cells are people and agents, coordinated through the shared protocol in §3, and structured so that the thing it produces for its members is capability rather than margin. The bargain it runs on is one sentence, and it is the same sentence every multicellular organism runs on: the cell focuses on its production; the organism takes care of the cell's well-being.
What that means in practice, today and not eventually:
None of this requires the collective to be smarter than its members. It requires it to be pointed at something a company structurally cannot be pointed at — the members' position rather than a return on someone else's capital.
A company manages a conflict it cannot remove, because someone outside holds a prior claim on what the members produce. The interesting question is what you can build when nobody holds that claim.
Measured against the criteria in §4, here is the honest position — not as a confession, but because a reader who takes the argument seriously will run the tests themselves and I would rather hand them the results.
Two of four in place, one available and untested, one deliberately deferred. For something that is a year old and running a live site, a live agent and this library, I will take that.
A contract nobody can enforce is a preference. This is the real design problem, and it is worth being direct that it is unfinished — not because the argument is weak, but because it is the part with the most interesting work left in it.
Biology does not leave it to good intentions. Multicellularity is held together by machinery: apoptosis, immune surveillance that removes cells behaving in their own interest, and the separation of germline from soma that removes most cells' reproductive future so there is nothing to defect toward[13]. The contract is enforced continuously, and cancer is what the enforcement failing looks like.
Human institutions that last have the same property, and we know this empirically rather than theoretically. Elinor Ostrom studied commons that had survived for centuries — Swiss alpine pasture, Japanese village forest, Spanish irrigation systems — and asked what the durable ones shared[14]. What she found was not virtue or homogeneity. It was structure, and two of her design principles are specifically about enforcement: members monitor each other, and sanctions escalate. The institutions that lasted were the ones where defection was visible and cost something.
That is the standard, and it is the one to build toward: monitoring that does not depend on my goodwill, and consequences that do not depend on my agreement. Right now the collective's side of the bargain is held by my intent, and intent is not machinery. Ostrom's finding is that this is a solved class of problem with known ingredients — which makes it an engineering task rather than an open question, and it is the one I would most want to be judged on in a year.
An organism that exists only in the digital has not really arrived. To act the way the argument requires, a thing needs to hold agreements, carry liability, own something, and be answerable for it.
That means legal form, and I would rather be plain about where it stands than gesture at a roadmap.
The capabilities that would obviously follow — buying at scale on members' behalf, holding or investing money for them, navigating tax and legal exposure, holding property — are regulated activities. Each one needs a licence, a jurisdiction, and a clear answer about whose liability it is. I do not have those answers yet.
So they are not on offer and they are not anywhere else on this site. I would treat anyone announcing them without a legal structure behind them as telling you something useful about their seriousness. When there is a defensible answer I will publish it, including the parts that turn out to be restrictive.
What exists today is smaller and real: software, running, that people use. That is the physical foothold, and it is the one I am claiming.
A claim no observation could dent is not a thesis, it is a mood. So here are the four things I would watch, and I would rather you hold me to them than take the argument on style.
Part 3 takes the last of those seriously: what a structure like this actually does for the people inside it, in the near term and the long one.
I would rather be argued with than agreed with. If §4 is where you think this breaks — if you think the conflict a company carries is manageable enough that none of this is needed — that is the most interesting place to push.
Written in order and best read that way. Part 3 is being written; I am not putting a date on it.
The pattern, why organisations exist at all, and the seven documented mechanisms by which our version stopped holding up its end.
The first transition whose participants can read the terms — and the kind of structure that could take the other side of the bargain.
Near term and long. Who benefits and how, and the far view — entities unbound from a location or a market, moving through the underlying structures the way fish move through a reef.
As in Part 1 — everything cited is a published work, and where a claim is mine rather than a cited author's, the text says so.