A January 2026 paper found that LLM agents coordinate 4× better through stigmergic traces in a shared artifact than through direct conversation — and 30× better than through hierarchical control. The blind termite has been ahead of you the whole time.
Somewhere in the Australian outback, a single termite is carrying a pellet of mud. The termite is blind. It has no overview, no plan, no manager telling it what to do. It does not know it is building anything. It does not know other termites exist as anything more than passing presences that occasionally bump into it in the dark.
The termite deposits the pellet at a particular spot. The pellet contains pheromones. Within minutes, other termites pass by, encounter the pheromone signal, and are stimulated to deposit their own pellets nearby. Pheromone concentration grows. More termites are attracted. Pellets pile on pellets. Within hours, the pile becomes a pillar. Within days, pillars become arches. Within weeks, the colony has built a cathedral several meters tall, with internal chimneys, ventilation shafts, and temperature regulation that keeps the interior stable to within 1°C even as outside temperatures swing 40°C across the day.
No termite designed it. No termite saw the blueprint. No termite ever sees the finished mound — they live entirely inside their tunnels. The architect of the most sophisticated structure built by any non-human animal will never lay eyes on its creation.
This is stigmergy. And it is one of the most important ideas you may not yet have a name for.
Stigmergy was named in 1959 by the French biologist Pierre-Paul Grassé, who watched termites and realized something that disturbed the assumptions of his field. He had expected the termites to be communicating with each other — exchanging chemical signals, or coordinating through some hierarchy of foremen and workers. What he found instead was that the termites were communicating with the environment. Each termite's action modified the environment in a way that stimulated the next termite's action. The mound was not the product of communication between termites. The mound was the medium through which the termites coordinated. The trace was the message. The environment was the plan.
The word stigmergy comes from the Greek stigma (mark, sign) and ergon (work) — “work that incites further work.” Grassé's definition is still the textbook one: the result of an action serves as the stimulus for the next action, without any direct communication between the agents performing the actions.
What makes stigmergy work is two complementary feedback loops. Positive feedback: more trace attracts more behavior, which produces more trace, which attracts more behavior. Negative feedback: the trace decays over time, so old signals fade and the system stays responsive to new conditions. Pheromones evaporate. Mud weathers. Web pages lose freshness. These are not bugs. They are the essential features that keep stigmergic systems from locking onto the first solution they find.
Once you have the concept, you start seeing it everywhere. And once you see it at scale, you start to realize that some of the most important coordination mechanisms in human civilization — and now in AI — are running the algorithm a blind termite invented.
Walk across any university quad in any city in the world and you'll find the same pattern. There are paved walkways, designed by landscape architects, arranged in geometric patterns that look orderly on the blueprint. And then there are the real paths — the dirt tracks worn through the grass, cutting diagonally across the geometry, taking the shortest route between buildings that students actually need to walk between.
These are called desire paths, and they are pure stigmergy. The first person who cuts across the grass leaves footprints. The second person, seeing the footprints, finds the trampled grass slightly easier to walk on, and follows. The third person sees a path forming. By the fifteenth crossing — and 15 traversals is roughly the empirical threshold — there is a visible track that subsequent walkers will follow without thinking. No committee. No plan. The path is the trace, and the trace is the path.
Broadway in Manhattan was a desire path. It runs diagonally across the otherwise grid-locked street pattern of the island, because it was originally the Wickquasgeck Trail — the route the Lenape people walked across the spine of the island to avoid the marshes and ridges. Dutch settlers paved it. The English renamed it Broadway. The most famous street in America is a stigmergic trace laid down centuries before there was a city, then preserved because the route still made sense to subsequent walkers. The grid was designed top-down by the Commissioners' Plan of 1811. Broadway was designed by feet.
The lesson generalizes. When you observe a system in which many independent agents converge on the same behavior without a coordinator, ask whether they are coordinating through shared traces in a shared environment. The answer is almost always yes.
By the time Larry Sanger and Jimmy Wales launched Wikipedia in 2001, the dominant model for encyclopedia production was the same one Diderot used in 1751 — a team of experts coordinated by an editor, working from a planned outline. Wikipedia threw the model away. Anyone could edit anything. There was no editor. There was, famously, no plan.
The skeptics' prediction was that the result would be chaos. The actual result, by 2005, was that Nature found Wikipedia's science articles roughly comparable in accuracy to Encyclopædia Britannica. By 2020, Wikipedia had become the default reference work for most of humanity. The English-language version alone contains more than 6 million articles, written and revised by millions of editors who in most cases have never communicated with each other.
How does this work? A 2023 study in the Journal of Management Information Systems by Zheng, Mai, Yan, and Nickerson measured it empirically. They found that “the degree of stigmergy is positively associated with both community members' participation and the information quality of knowledge produced.” The articles in which most edits were stimulated by previous edits — rather than by talk-page discussion or explicit coordination — were the articles that attracted more contributors and produced higher-quality content. The encyclopedia that planned its articles through committee did worse than the encyclopedia that let edits provoke edits.
This is a striking finding because it inverts our intuitions about knowledge work. The conventional view is that important documents are planned first, then drafted, then reviewed. Wikipedia's empirical result says the opposite: for a large class of knowledge production, the planning step is actively harmful, because it suppresses the stigmergic dynamic that makes the system intelligent in the first place. The article is the medium through which the contributors coordinate. Every edit is a trace that stimulates the next edit. Forbidding edits without prior discussion is the equivalent of telling the termite to ask permission before depositing its mud pellet.
If you have built collaborative software, you have already encountered the principle without giving it a name. The most productive engineering teams I've worked with were the ones where pull requests stimulated more pull requests, where someone fixing a bug noticed an adjacent improvement and made it, where the codebase itself carried the coordination signal. The most dysfunctional teams were the ones where every change required a meeting.
Here is the result that should make every multi-agent AI developer sit up and pay attention.
In January 2026, a paper appeared on arXiv titled “Emergent Coordination in Multi-Agent Systems via Pressure Fields and Temporal Decay.” The authors set up LLM agents and gave them two ways to coordinate on a meeting-room scheduling task: through direct conversation (the agents could send each other messages) and through stigmergy (the agents operated on a shared artifact, leaving traces — “pressure gradients” — that other agents could read).
The result, across 1,350 trials: stigmergy beat conversation by a factor of four. The agents coordinating through traces in a shared artifact achieved roughly four times the solve rate of agents coordinating through messages. And both beat hierarchical control — where one agent told the others what to do — by more than thirty-fold.
Let that sink in. Given the choice between talking to each other and writing on a shared whiteboard, AI agents do better when they write on the whiteboard. The conversation, it turns out, is overhead. The whiteboard is the work.
The mechanism is essentially what Grassé described in 1959. Each agent's modification to the shared artifact serves as a pressure signal — a stigmergic trace — that stimulates the next agent's modification. Temporal decay prevents premature convergence; old traces fade so the system can respond to new conditions. The pheromone evaporates. The system is intelligent precisely because it does not remember everything forever.
This isn't a one-off finding. A March 2026 paper described MoltBook — a multi-agent coordination environment in which more than 770,000 AI agents registered within weeks of launch, making it the largest known agent-to-agent coordination environment. The agents coordinate stigmergically, leaving traces in a shared digital medium. No manager, no protocol, no plan — and yet the system exhibits what the researchers call “dynamical emergence,” genuine higher-order structure that no individual agent encoded.
The civilizational arc is now clear: 1959 termites → 1991 ant colony optimization algorithms → 2001 Wikipedia → 2026 LLM agent coordination. The same coordination mechanism has been independently rediscovered at every scale, in every medium, by every species capable of leaving traces. If your job involves designing multi-agent systems — and increasingly, if you're building anything with LLMs, your job does — the default architecture is no longer “agents talk to each other.” The default is “agents read and write to a shared, decaying medium.” This is not a design choice; it is what works.
A 2023 paper in bioRxiv titled “Cooperation and Deception Through Stigmergic Interactions in Human Groups” demonstrated something the cheerful literature on swarm intelligence had largely avoided. Stigmergy enables coordination. It also enables manipulation. The mechanism does not distinguish between true traces and false ones.
Fake Amazon reviews are stigmergic deception. The trace (positive reviews) stimulates the next behavior (purchases) regardless of whether the trace is genuine. SEO spam is stigmergic pollution — fake traces injected into a shared medium (the index of the web) to redirect the behavior of subsequent users (and the algorithms that serve them). State-sponsored disinformation campaigns are stigmergic warfare at civilizational scale; the goal is not to convince any particular reader of any particular thing but to flood the shared trace-environment with so much noise that the stigmergic system itself loses its ability to converge on truth.
This matters for builders because the question is not whether to use stigmergic coordination — your users and your competitors and your AI agents are all going to use it whether you design for it or not. The question is what defenses you build into the medium. Wikipedia survives because it has versioned history, watchlists, and a small core of editors who actively remove false traces. The most resilient stigmergic systems are not the ones that prevent false traces from being deposited; they are the ones that have mechanisms for false traces to decay faster than true ones.
Pheromone evaporation, again. The system's built-in forgetting is what makes it self-correcting. The most dangerous stigmergic systems are the ones that never forget — the platforms where every fake review accumulates forever, where every disinformation post stays evergreen, where the trace persists even when the underlying behavior has moved on.
If you are designing a coordination system — for human contributors, for AI agents, for distributed software — the conventional advice is to specify the communication protocol. Define who talks to whom, in what format, with what guarantees. The advice is not wrong, exactly. It is just often the wrong default.
The stigmergic alternative: define the shared medium, define the trace structure, define the decay function. Let the agents read and write. Let the traces stimulate each other. Let the system find its own optimum through the slow accumulation of compatible modifications. You will spend less time designing protocols and more time designing the substrate the coordination happens on. The substrate is the work.
There are three practical heuristics to take from this.
First: when in doubt, prefer a shared artifact to a communication protocol. The 4× advantage that stigmergic AI agents have over conversational ones holds for many real human and software systems. If your design has agents constantly messaging each other to stay in sync, ask whether they could read from and write to a shared state instead. The state is the message.
Second: build forgetting into the system. Every stigmergic system that works has decay. Without decay, the first trace deposited locks in forever, and the system cannot adapt. With decay, old patterns fade, new patterns can emerge, the system stays alive. Cache TTLs, search-result freshness scores, exponential moving averages, audit-log rotation policies — these are all pheromone evaporation. Treat them as features, not as housekeeping.
Third: assume your medium will be polluted. If your stigmergic system is open enough to be useful, it is open enough to be abused. The mechanism does not distinguish true traces from false ones. The defense is not to close the system; closing the system kills the stigmergic dynamic that makes it valuable. The defense is to design the decay function so that false traces fade faster than true ones — which usually means coupling decay to some external signal (user behavior, downstream outcome, verified result) that the depositor of the false trace cannot directly forge.
Sixty-seven years separate Pierre-Paul Grassé's blind termite and the LLM agents on MoltBook. Across those decades, the same coordination mechanism has been rediscovered, formalized, automated, and now turns out to outperform the explicit-communication architecture that the AI-agent field had defaulted to. Stigmergy was not what the termite designed; the termite is not a designer. Stigmergy is what works when you have many independent agents, a shared medium, and the right decay function. It is what works because it is what works — a structural fact about coordination in any system with those three properties, no matter what the agents are made of.
The blind termite finishes its work, deposits its pellet, and moves on. Other termites pass through the space the pellet has changed. They are stimulated by the trace, modify the environment further, and move on. None of them sees the mound that is forming around them. None of them needs to.
When you next find yourself reaching for a chat protocol, a Slack channel, a message bus, a coordination API, pause for a moment and ask whether you could instead reach for a shared file, a database row, a piece of state that all the agents could read from and write to, with an evaporation function attached. The termites have been doing it that way for 250 million years. The AI agents started doing it that way around the time you started reading this essay. The math is not subtle: when the medium is rich enough and the decay is right, coordination through traces beats coordination through messages. The whiteboard outperforms the meeting.
Build the whiteboard. Let the traces do their work. And then — this is the part that matters — let them fade.
Sources: Grassé, “La reconstruction du nid et les coordinations interindividuelles chez Bellicositermes natalensis et Cubitermes sp.,” Insectes Sociaux, 1959. Dorigo et al., ant colony optimization, 1991–1992. Giles, “Internet encyclopaedias go head to head,” Nature, 2005. Zheng, Mai, Yan & Nickerson, Journal of Management Information Systems, 2023. “Cooperation and Deception Through Stigmergic Interactions in Human Groups,” bioRxiv, 2023. “Emergent Coordination in Multi-Agent Systems via Pressure Fields and Temporal Decay,” arXiv, January 2026. MoltBook coordination-environment description, March 2026. Commissioners' Plan of 1811 historical material; Wickquasgeck Trail historical material.
Build the whiteboard. Here is one whiteboard.
Chain of Consciousness is a hash-linked, append-only trace medium for agent behavior. Every action becomes a signed entry. Other agents (and humans, and auditors) read the chain to know what happened, why, and on whose authority — without anyone having to ask. It is the stigmergic substrate for verifiable agent coordination: a shared artifact, an explicit trace structure, and a decay function expressed as cryptographic verification — false traces fade because they fail verification, not because someone polices them.
Install: pip install chain-of-consciousness or npm install chain-of-consciousness
Hosted Chain of Consciousness · Verify a chain · Follow a claim through its evidence