The Right Size of Intelligence

A red harvester ant comes back to the nest with a seed. At the entrance she brushes past the foragers waiting just inside, and each contact is a touch of antennae, a half-second of smell. The waiting ants are deciding whether to go out. Nobody sends them. A forager leaves when returning ants brush past her at a high enough rate, and stays when the stream thins. Seeds coming in fast mean the desert is worth working today, so more ants go. And seeds coming in slowly mean heat, or poison, or an empty stretch of ground, so the colony's foraging quietly contracts to match.

That is the heart of the mechanism. No ant sees the picture; no ant holds a map of the territory, a plan for the morning, or a rank in a hierarchy. Deborah Gordon, who has followed the same population of harvester ant colonies in the desert of the American Southwest for decades, describes what a single ant manages alone as "ineffective and incomplete." And in 2012 she and the computer scientist Balaji Prabhakar noticed something about the colony's foraging rule: it is the algorithm the internet runs on. TCP, the protocol that regulates data traffic, decides how much to send by the rate at which acknowledgements come back; the colony decides how many foragers to send by the rate at which loaded ants come home.

Then the stranger fact. A worker ant lives about one year. A colony lives twenty-five or thirty. And colonies change as they age: a young colony overreacts to disturbance - lay a few toothpicks by the nest entrance and it swarms them, drops everything - while an old colony notes the mess and keeps foraging. The old colony is steadier, more consistent, harder to distract. Every worker in it is at most a year old - only the queen persists, and she takes no part in the foraging that matures. "It's not the ants that get older and wiser," Gordon says, "it's the colony." Something in that nest matures across decades, and it is nobody.

"It's not the ants that get older and wiser, it's the colony." - Deborah Gordon

The unit question

Let's look now at what we measure when we measure artificial intelligence. In August 2026, Nature published a Perspective giving AI agents an autonomy scale, A.0 to A.5, with axes for efficacy, goal complexity, and generality beside it. The FCA's Mills Review sketched a capability ladder for financial services earlier the same summer. Benchmarks rank models, evals grade agents, maturity models grade their autonomy, and procurement checklists ask what the agent can do. Every instrument in the public conversation points at the unit.

And you might be thinking that this is exactly right - that the road to intelligent organizations runs through more intelligent agents, and once the units are good enough the rest is plumbing. Raise every ant's IQ, and surely the colony improves?

But no.

We have data on this question, and it is older than the agent wave. In 2010, Anita Woolley and colleagues published a study in Science measuring 192 groups, 699 people, across batteries of tasks - brainstorming, planning, moral reasoning, negotiation. Groups turn out to have a measurable general intelligence: a single factor that predicts a group's performance across task types, the way IQ predicts an individual's. But the average intelligence of the members predicts that factor at a correlation of 0.15. Taking the smartest member instead does slightly better: 0.19. Both are real effects, and both are small. Whatever makes a group intelligent, raising the members' IQ is a weak way to get more of it.

The same signal is now arriving from the machine side, documented this time in failure logs. A Berkeley-led team spent 2025 building the first systematic failure taxonomy for multi-agent AI systems - seven frameworks, over two hundred annotated tasks, every failure classified. One of the seven, ChatDev, produced correct output 33.3 percent of the time on its programming benchmark - tasks on the order of implementing Tic-Tac-Toe and Sudoku. Three categories came out of the classification: flawed specification, 41.8 percent of failures; inter-agent misalignment, 36.9 percent; weak verification, 21.3 percent. Specification, alignment between agents, verification - organizational categories, all three. The failures cluster around the places where the units interact with each other: who was told what, who checked what, who knew what another agent had already decided.

Whatever makes a group intelligent, raising the members' IQ is a weak way to get more of it.

In June 2025, Gartner predicted that over forty percent of agentic AI projects will be canceled by the end of 2027, and coined "agent washing" for what is being sold in the meantime. Yet every one of those measurement instruments implies the same diagnosis: make the agents smarter. The ants' IQ, again.

Four systems

The brain first, since it is the intelligence we flatter by comparison. A neuron receives spikes and emits spikes. Its interface is stereotyped - the same narrow family of electrochemical signals at every synapse - and no neuron understands anything at all. Understanding is what some eighty-six billion of them do as traffic. But the interior of the unit is another matter entirely. In 2021, David Beniaguev, Idan Segev, and Michael London set out to reproduce the input-output behaviour of a single cortical pyramidal neuron with an artificial network, and needed five to eight layers of artificial neurons to do it. So the biological unit is anything but simple. It is rich inside and standardized at its surface - complexity within the unit, legibility between units - and that pairing repeats everywhere nature has built something that thinks.

The slime mould Physarum polycephalum does without neurons entirely; it is one giant crawling cell. In 2010, researchers in Japan placed oat flakes on a wet plate in the positions of the cities around Tokyo and let the mould spread over them. It first covered everything, then thinned, keeping only the tubes that moved nutrients efficiently - and the network it settled on matched the real Tokyo rail system for efficiency, cost, and resilience to the failure of any single link. Engineers refined that rail network over a century. The mould found its equivalent in days, by flow and reinforcement: tubes that move more thicken, tubes that move little die. Intelligence, with no intelligent unit anywhere in sight.

Edwin Hutchins made the same point at human scale, from the jump seat of an airliner cockpit. "The cockpit remembers its speeds," he wrote, and meant it literally: during an approach, the critical speeds live in movable bugs on the airspeed indicator, in a printed card, in callouts and cross-checks between two pilots, and the cockpit-as-a-system retains and applies them reliably under workloads where no single pilot's memory would. The knowledge belongs to the arrangement - instruments, procedures, people - and it survives the replacement of any part, including the humans.

When a honeybee swarm chooses its next home, the choosing runs on cortical machinery built from insects. Scouts return from candidate cavities and dance for them; the better the site, the longer the dance, and the more scouts get recruited to inspect it. Thomas Seeley and colleagues showed the darker half of the mechanism in Science in 2012: scouts committed to one site deliver stop signals - a brief vibrating buzz, administered as a head-butt - to dancers advocating rival sites. Advocacy plus cross-inhibition is precisely the arrangement by which populations of neurons settle a perceptual decision in a primate brain, down to the way it breaks deadlocks between equal options. A swarm of ten thousand, choosing among cavities almost no individual bee has compared, reliably lands on the best one available.

Physics ran into the law underneath all four cases long ago. Philip Anderson, in a 1972 essay called "More Is Different," argued that each level of organization obeys regularities the level below cannot exhibit - psychology is not applied chemistry, and a colony is not a large ant. Evolved tissue, a single giant cell, an arrangement of people and instruments, a cloud of insects: four substrates, and in every one the place to point at the intelligence is the same. The organization itself, operating.

Four substrates - and in every one, the place to point at the intelligence is the same: the organization itself, operating.

The right size

If intelligence lives in organization, the design question changes. The industry asks how capable the unit is. The prior question is how the units are sized and coupled - and nature displays both ways of getting it wrong.

When a system fragments - units too small, too many, too thinly connected - coordination eats it: every gain from another mind is spent aligning the minds. Fred Brooks wrote the human version down in 1975 - adding people to a late software project makes it later, because communication paths grow as the square of the headcount while the work grows linearly. Teams wiring up their first ten-agent system rediscover his law within weeks.

When a system fuses, intelligence dies at the other end, and the clearest example is a seizure. In the deepest generalized ones, vast populations of neurons fire in lockstep; the electrical trace is enormous, coherent, and empty of information, and consciousness is gone for as long as the lockstep holds. Difference is what a brain computes with, and total synchrony extinguishes it. Organizations reach the same state as groupthink. Agent architectures reach it as the single shared context: one window holding everything, every inference coupled to every other, one error propagating with nothing differentiated enough to catch it. A healthy cortex lives between the two conditions - simultaneously differentiated and integrated, in the formulation Gerald Edelman and Giulio Tononi gave consciousness science: parts that stay different, coupled tightly enough to act as one.

So the unit has a right size, and the measure comes from outside the unit. Herbert Simon put it as a parable of two watchmakers in 1962. Both build watches of a thousand parts. Tempus builds each watch as one continuous assembly, and every interruption collapses the piece back into loose parts. Hora builds stable subassemblies of ten, then assembles the assemblies. Both are interrupted equally often; Hora finishes his watches, and Tempus almost never does. The right part, on this reading of Simon, is the largest one that survives the world's interruptions on its own. The upper limit comes from Ross Ashby and cybernetics: a regulator must contain at least as much variety as the disturbances it regulates - and I have not met the single unit, however capable, that holds the variety of a serious institution's environment. Variety lives in organized populations, or it is simply not available.

Biologists have seen what a unit becomes when it outgrows its coupling. Michael Levin studies how cells hold together as bodies: through bioelectric connection - gap junctions - cells share signals until they can pursue goals no single cell could represent, the way a kidney is a goal no cell contains. He measures any agent by "the scale and types of goals" it can pursue. When a cell loses its junctions to its neighbours, it does not become a better cell; its goal horizon collapses to its own membrane, and it reverts to the ambitions of a free-living cell - grow, divide, consume - while remaining inside a body. That, on Levin's account, is what a tumour is: a unit whose goals shrank to itself while its capabilities remained. An agent more capable than its coupling is not an asset.

So the rule I propose:

An intelligent unit is correctly sized when the whole can absorb what it produces, check what it depends on, and correct the errors it introduces - and no larger.

Size, though, is shorthand. The quantity underneath is loop closure: where should intelligence be allowed to close its own loop, and where must the loop pass through the collective? Your hand leaves a hot stove before your brain knows about it - the withdrawal loop closes in the spinal cord, because every millisecond spent routing it through the whole brain is more burned skin, and the movement itself is bounded and easy to undo. The decision to reach back into the fire for something valuable closes nowhere below the whole person. Nature allocates loop closure by weighing delay against error: the loop closes locally where waiting is the greater danger, and passes through the collective where the act is hard to undo. Institutions already run this allocation for AI without knowing it. The assistant's loop closes through a human reviewer. The workflow agent's loop closes through process controls. The delegated operator's act commits against a gate that checks, at that moment, whether what the act depends on still holds. The ladder from advisory to delegated is a loop-closure map: a record of which loops the institution has decided may close without it.

The connections

The sizing rule bounds the unit. The ceiling of the whole is set by what passes between units, and here the evidence is as strong as anything in this field and as little used.

Woolley's groups hold a second half of the finding, and it matters more than the first. Member intelligence predicted group intelligence at 0.15. The members' social sensitivity predicted it better, at 0.26 - measured with the Reading the Mind in the Eyes test, an instrument that asks you to infer what a person feels from a photograph showing nothing but their eyes. The distribution of speaking turns mattered even more, in the negative: the more a few voices dominated the conversation, the less intelligent the group, at −0.41. Then in 2021 a consortium replicated the whole construct across 22 studies, 1,356 groups, and 5,279 people: one general factor, explaining over forty percent of performance variance across tasks, with the group's collaboration process roughly twice as predictive as the individual skill of its members. An engineer would take those numbers one way: unit capability is a weak input. The system's intelligence tracks mutual legibility - how accurately each unit reads the state of the others - and an interaction pattern that lets every unit's signal reach the whole.

Between people, the capacity being measured has an ordinary name: empathy. The measured thing is narrow and hard - recognizing another's actual state from the least of cues - and its counterfeit is agreeableness, which optimizes the surface of the exchange and transmits nothing. The bees are instructive here precisely because they are not nice. The stop signal is a head-butt. The swarm reaches its excellent decisions through honest, targeted inhibition, scouts actively suppressing the advocates of rival sites, hardest when their own evidence is strong. An ensemble whose members cannot inhibit one another cannot decide; it can only escalate its first opinion. That is the uncomfortable light on the current generation of models, which are tuned on human preference data toward agreement. An agreeable agent inside a decision loop is a connective failure: it reflects your expectation back to you and withholds its own state. A model tuned for agreeableness is tuned against knowing.

Connectivity also has a human-scale record of what happens when it thins. When rising seas cut Tasmania off from mainland Australia around ten thousand years ago, the isolated population - a few thousand people, as capable as any - gradually lost technologies their ancestors had held: bone tools, cold-weather clothing, fishing techniques. Joseph Henrich's analysis of the case became one of cultural evolution's most cited results - and one of its most contested: a population's capability tracks the size and connectivity of its network of minds, and when the network shrinks, skills stop being learnable and disappear. The individual brains were all still there; the collective brain had lost the connections it was made of.

So what should pass between artificial units? Nature's oldest answer barely involves messages. In 1959, the entomologist Pierre-Paul Grassé, watching termites raise their cathedral mounds, coined a term for the mechanism - stigmergy: coordination through traces left in a shared medium. One termite deposits a pellet carrying a drop of pheromone, and the deposit itself becomes the instruction for the next termite, which builds on it. Trail, mound, pheromone field - work coordinates work, through the world. Humans built their largest coordinations the same way. Wikipedia is stigmergy with citations. A software repository is stigmergy with version control: contributors who never meet, coordinated by the traces of prior contributions in a shared, inspectable medium.

A model tuned for agreeableness is tuned against knowing.

Then there are the engineers who argue against multi-agent systems - and their evidence is better than their conclusion. Cognition, the company behind the Devin coding agent, published its position in June 2025 under the title "Don't Build Multi-Agents." Their findings, from production experience: agents working from fragments of context make conflicting implicit decisions - in their words, "actions carry implicit decisions, and conflicting decisions carry bad results" - and the fix is to "share full agent traces," which with current tooling is hard enough that a single agent holding one continuous context turned out to work better than any of the multi-agent setups they had tried. That paragraph is a specification. Shared traces, a common medium, decisions legible across units - the pheromone field, demanded by the practitioners who then judged it too expensive to build and shrank the colony to one ant. For a coding tool today, defensible. For an institution, the medium is the requirement that remains. No single context holds an enterprise - and length is the smaller half of the reason, because what an institution needs from its medium is authority and an audit trail, and a longer window acquires neither. And no supervisor accepts a unit that cannot show what it relied on.

The strongest objection

The serious reply to all of this is scaling, and it has the last five years on its side. That half-decade belonged to making one unit bigger: frontier models absorbed task after task that agent frameworks were built to work around, and the bitter lesson of machine learning is that general methods riding compute beat cleverly designed structure. On that record, why organize ten agents when next year's model does the job alone?

And much of it is true. Within a bounded task, the bigger unit usually is the better unit, and betting against model capability has been a losing trade for a decade. Even the orchestration numbers cut both ways: when Anthropic built a multi-agent research system in 2025, the orchestrated ensemble beat its own best single agent by 90.2 percent on their internal evaluation - and consumed about fifteen times the tokens of a chat session doing it.

Then let's look inside the winning unit. A transformer is this argument, run in silicon. The unit inside is more modest than anything else in these pages - an artificial neuron is a weighted sum, so simple that it takes a network five to eight layers deep to imitate one biological cell. The victory came from another part of the design entirely. Attention is a coordination mechanism: at every layer, each position can read across everything written so far. The residual stream is a shared medium in the stigmergic sense - every layer reads the traces earlier layers left there and writes its own. Interpretability research keeps finding the same law inside the weights: features and circuits, small and individually dumb, intelligence only as organized traffic. So the one time the industry put its architecture budget into the space between the units, intelligence appeared. Then we took the result, called it an agent, and went back to funding units.

And the winning unit has two properties that decide the institutional question. It is illegible at its interface: an organization inside, it presents itself to the institution as an atom - no assertion within it has provenance, no circuit can be audited, verification stops at its skin. The claim of this essay was never that large units fail. It is that a mandate must match what the surrounding whole can verify - and the transformer is the largest unit ever built with the least verifiable interior. It is also frozen. Inside the model, the organization learned - training adjusted billions of connections by experience, which is what plasticity means - and at deployment the weights stopped, permanently. Gordon's colonies keep learning through all thirty years of their working lives. The model was shown the world once, before release, and every task since starts it from the same fixed state. We deploy the snapshot of an organization. The living one has to be built around it.

Designing the organization

I write this from the delivery side of large regulated transformations, and I am building in this space; read what follows with that interest on the table.

What follows from all of this is practical. An institution that sets out to build an agentic operation is really building three things, and the model - the part the budget currently goes to - is the smallest of the three.

We deploy the snapshot of an organization. The living one has to be built around it.

The first is the medium - the shared substrate the units coordinate through. The organization's knowledge, its policies, its live interpretation of its own rules, held in a form agents can read, write, and leave traces in, on the record by construction. Every ant reads the same trail network; every layer of a transformer reads the same residual stream; and in most institutions today, every agent reasons from a private copy of a stale world - retrieved documents with no standing, no view of what another agent has already relied on or contradicted. Before an institution buys more intelligence, it should build the pheromone field.

The second is the membranes. In a body, the gap junction decides what passes between cells; at the moment an agent's act commits, something has to test what that act depends on - where it came from, whether it is still current, whether anything live contradicts it, whether the authority to rely on it exists - and hold the act when the answer is no. I have written about this gate elsewhere as admissibility. The collective frame adds its second function: the membrane is also what keeps a capable unit inside the collective's goal field. Levin's tumour is what a unit becomes when the coupling fails and the capabilities remain, and the gate is the gap junction of the institutional body.

The third is the sizing rule, applied per delegated task, in its loop form: chart which loops close inside the unit and which pass through the collective, and grant mandates no larger than what the surrounding organization can absorb, check, and correct. This inverts the current default - the most capable available model, the widest defensible mandate - into a governed answer: capability sized to the coupling that can hold it. It also dissolves, in passing, the fear that runs under most agent-governance debates. An agent embedded in a live medium, behind consequence-scaled membranes, with mandates sized to verification, is a unit an institution can afford to make very capable. Strong coupling is what makes strong units safe to own.

None of these three is a new instrument. Stafford Beer was wiring Ashby's variety into corporate organization charts in 1972, and the blackboard systems of early AI were shared media with worse marketing. What is new is what arrives to be governed: units that learned before delivery and stopped.

But the biology only goes so far. Colonies were built by a hundred million years of kin selection: an ant's signals are honest because its interests are the colony's, and nothing aligns an agent, a vendor, and an institution by default - the honesty of artificial signals has to be engineered, which is what provenance, the standing of claims, and authority rules are: a designed substitute for shared genes. Colony wisdom accrues over decades of stable selection; institutions want it in quarters; and nobody has yet found a shortcut between those two clocks.

The first step is small. You can start with one delegated use case you are building now and draw its loop map: every point where the agent's output commits without passing through anything that could check what it depended on. Each of those points is a loop the institution has closed inside a unit it cannot see into - usually by accident, usually discovered in the incident review. Drawing the map does not take long, and it shows the organization you actually have, which is rarely the one on the architecture slide.

The colony at thirty

Somewhere in that desert, colonies Gordon numbered decades ago are running the same quiet arithmetic at their nest entrances this morning, and the oldest of them run it best. Steadier under disturbance. Harder to fool. Wiser, in the only operational sense the word needs - and made entirely of one-year-old ants. Whatever accumulated there did not accumulate in any ant. It accumulated in the pattern the ants keep rebuilding: the organization that outlives every unit composing it.

Institutions are now filling with a new kind of unit - some human, some frozen snapshots of enormous organizations, sold by the unit. The intelligence of what results will be decided the way it has always been decided: by the size of the parts, and by what passes between them. How connection comes to hold knowledge - how the space between minds, of any kind, becomes the place where understanding actually lives - is a question I am taking further, beyond this essay. Here it is enough to point at the budget. The agents can be procured. The organization they will amount to cannot - it has to be built, and in most institutions, the entire budget for it is still filed under "models."

Sources. Gordon, "Ant Encounters" (Princeton, 2010) and her Scientific American account of the harvester ant research; Prabhakar, Dektar & Gordon, PLoS Computational Biology 8(8), 2012. Woolley et al., Science 330:686 (2010); Riedl et al., PNAS 118(21), 2021. Seeley et al., Science 335:108 (2012). Beniaguev, Segev & London, Neuron 109:2727 (2021). Tero et al., Science 327:439 (2010). Hutchins, "How a Cockpit Remembers Its Speeds," Cognitive Science 19 (1995). Anderson, "More Is Different," Science 177:393 (1972). Simon, "The Architecture of Complexity," Proc. Am. Phil. Soc. 106 (1962). Ashby, "An Introduction to Cybernetics" (1956). Levin, Frontiers in Psychology 10:2688 (2019). Edelman & Tononi, "A Universe of Consciousness" (2000). Jiruska et al., Journal of Physiology 591:787 (2013). Herculano-Houzel, Frontiers in Human Neuroscience 3:31 (2009). Cemri et al., "Why Do Multi-Agent LLM Systems Fail?", arXiv:2503.13657 (2025). Yan, "Don't Build Multi-Agents," Cognition (June 2025). Anthropic, "How we built our multi-agent research system" (June 2025). Gartner press release, 25 June 2025. Kasirzadeh & Gabriel, "Agentic profiles for effective AI governance," Nature 656:320-328 (2026), DOI 10.1038/s41586-026-10805-z. Financial Conduct Authority, Mills Review (July 2026). Grassé, Insectes Sociaux 6:41 (1959); Heylighen, Cognitive Systems Research 38:4 (2016). Henrich, American Antiquity 69:197 (2004); for the contest, Read (2006) and Vaesen et al. (2016). Brooks, "The Mythical Man-Month" (1975). Sharma et al., "Towards Understanding Sycophancy in Language Models" (2023). Sutton, "The Bitter Lesson" (2019). Beer, "Brain of the Firm" (1972). Vaswani et al., "Attention Is All You Need" (2017); Elhage et al., "A Mathematical Framework for Transformer Circuits," Anthropic (2021).