The mound with no architect
A termite colony builds structures that no termite has seen. Bellicositermes workers raise arches, chimneys and ventilation columns metres tall relative to their own size, and the finished form obeys no plan held anywhere in any single insect's nervous system. Each worker carries a soil pellet, moistens it with a pheromone-bearing secretion, and drops it where the scent of previous deposits is strongest. The pellet is a mark. The mark is also a message, but not one sent to anyone. It sits in the world, and the next worker to pass reads it the way a hiker reads a trodden path: not as an instruction but as a fact about where others have already been. Enough pellets accumulate at one point to form a pillar; two pillars rising independently curve towards each other because each is reading the accumulating scent of the other, and they meet as an arch.
Two properties make this more than a curiosity about insects. First, the state of the task lives in the environment, not in any agent. No termite holds a plan; the mound holds the plan, distributed across its own surface as a pattern of chemical concentration. Second, the marks decay. Pheromone evaporates over a timescale of minutes to hours. A pellet dropped an hour ago recruits less strongly than one dropped a minute ago, and a pellet from a source that has stopped producing eventually recruits nothing at all. This second property is easy to treat as an engineering nuisance — surely a permanent mark would work better than a fading one — and that reading is exactly wrong. The decay is what lets the system track a world that keeps moving. Without it, an abandoned route would remain as attractive as an active one forever, and the mound would fossilise around its own history rather than continuing to respond to its present.
Coordination achieved this way needs no message passing, no shared plan, no roster of who is currently working. It needs only that the environment be legible and that old marks fade at roughly the rate the underlying situation changes. This is the general mechanism, and termites are only its first documented case.
Grassé's term, and its second life
Pierre-Paul Grassé named it in 1959. He was a French entomologist trying to explain termite construction without invoking a supervisor, a blueprint, or communication between individual workers, none of which the anatomy or behaviour of the insects supported. He coined stigmergie from the Greek stigma, a mark, and ergon, work: the mark stimulates the work. His formulation was precise and stayed largely within entomology for two decades.
It resurfaced in the 1980s and 1990s when swarm robotics and combinatorial optimisation needed a control scheme that did not require every agent to know about every other agent. Marco Dorigo's ant colony optimisation made the mechanism computational: virtual ants deposit virtual pheromone on the edges of a graph while searching for short paths, the pheromone evaporates at a chosen rate, and the rate itself becomes a parameter an engineer can tune. Turning decay into a dial, rather than treating it as biological incidental, was the move that let stigmergy travel out of biology and into anything with a trace field and a clock.
Where the axis is pointing
The lineage from Large Language Model to Large World Model to Large Universe Model is usually described along an axis called intake: how much of the world a system takes in, and how current that intake is. Stated that way it sounds like a matter of degree — more data, more often. Stigmergy suggests it is something sharper: a control regime that either holds or fails to hold, and intake is the condition for holding it.
A Large Language Model coordinates its outputs against a trace: the corpus it was trained on. But that corpus stopped changing at some cutoff. It is a termite reading a photograph of last year's mound. The deposits are legible — the shapes are all there — but nothing the system does adds a new mark, and nothing that has happened since is reflected in what it reads. It has a trace with the decay term surgically removed, because there is no more depositing to decay away from. A Large World Model does better: it senses a live scene, and it reads that scene while it is present, the way a worker reads pheromone concentration at the pillar it is standing next to. But when the episode ends, the trace is gone with it, and the next episode starts fresh. Decay is restored, but only inside one bounded window.
The Large Universe Model is the case where every stream that was ever running is still running, each one writing marks that are read, acted on, and rewritten, with the age and source of each mark carried alongside it rather than discarded on ingestion. That last part is not decoration. Provenance is what lets a system tell a fresh mark from a stale one without needing to have watched the whole history — the equivalent of the mound knowing not just that pheromone is present but roughly when it was laid.
The reasoning that gets from stigmergy to a terminal rung is this: if control is achieved by reading traces in a shared, changing environment, then freshness of read is not a nice-to-have, it is the condition under which reading means anything. A stale trace does not degrade gracefully. It recruits with full confidence towards a target that has already moved — a worker still building at a spot where the food source is long exhausted. A frozen corpus is a trace stripped of its decay term. A bounded scene restores decay but only within one episode, then discards even that. Continuous intake across everything still running, tagged so beliefs can be aged, is what stigmergic control actually requires to function, and there is no further category past "everything, still arriving, dated." What remains open past that point is quantity, trust, and retention — real questions, but not this one.
The misreading to disown
The common mistake is to hear "stigmergy" and conclude that centralised, planned control is unnecessary — that intelligence should therefore be swarm-like, decentralised, emergent from many simple parts. This inverts what the concept establishes. Stigmergy is a claim about where state lives, not about how many agents there are or how clever each one is. A single sophisticated reasoner, alone, coordinating its own successive actions by continuously reading a changing environment, is fully stigmergic. The argument here concerns intake and decay, not population size or agent simplicity. Treating it as an anti-hierarchy thesis keeps the biology and throws away the control theory that made the biology instructive.
Three objections, taken seriously
Termites succeed because they are simple and numerous. Sophisticated reasoners don't need to offload state into the world — importing an insect scheme mistakes a workaround for scarce cognition as a principle for abundant cognition.
Right about the origin, too narrow about the scope. Stigmergy is not confined to simple agents; it is the only scheme that scales when the set of participants is open and unknown at design time. Version control, market pricing, and immune surveillance all run this way, and none of the participants is a termite. Direct messaging costs grow with the square of the number of participants and require knowing who they are in advance. Trace-reading costs one write and admits anyone who arrives later. Sophistication does not remove this constraint — it raises the cost of ignoring it.
Most knowledge doesn't evaporate. Arithmetic and grammar are stable for decades. A frozen corpus handles this fine; continuous intake buys freshness only on a thin volatile margin.
This one genuinely narrows the claim. The stable core is real, and a snapshot serves it well. But nothing inside a snapshot marks which parts are stable. Half-lives are visible only by watching, over time, that a road closure decays in hours and a coastline in millennia. A system with continuous intake and provenance can measure and price that difference belief by belief. A frozen system must assume one decay rate for everything, which is wrong for almost everything in it. The gap is not freshness alone; it is the ability to know what needs freshness at all.
Continuous trace-reading is exactly what produces coordination pathologies — ants locked into bad trails, the Flash Crash erasing roughly a trillion dollars of notional value in thirty-six minutes as automated readers reacted to each other's marks. Freezing intake is a circuit breaker, not just a limitation.
The pathology and its cause are both correctly described. But every working stigmergic system answers it with a damping term inside the trace layer itself — pheromone evaporation, trading halts, exponential backoff, quorum thresholds — and every one of those remedies requires continuous reading to operate. You cannot damp a signal you have stopped receiving. Freezing intake does not stabilise the loop. It removes the system from the loop, and from any capacity to help stabilise it.
What this does and does not establish
Stigmergy establishes that continuous, decay-aware intake is the condition for a certain kind of coordination to work at all, and that a frozen or bounded trace is not a weaker version of that condition but a different, incomplete one. It does not establish that more streams or more agents make a system smarter, that decentralisation is preferable to hierarchy, or that the Large Universe Model, as an argued category, already exists as something built and running. It says where the ladder's top rung is on one specific axis. It says nothing about how to climb the others.