BONABEAU APPS
Stigmergy

Work leaves traces.
Traces direct more work.

Stigmergy is how a colony of termites builds a cathedral while every termite behaves as if it were alone. The nest itself is the coordinator.

From Greek stigma (a mark, a sting) and ergon (work). Coined by Pierre-Paul Grassé in 1959 to explain how termites rebuild their nests.

Pillars forming

Each dot is a termite dropping pellets. Fresh pellets carry a pheromone that attracts more pellets, then fades. Drag the builder slider below about 15 and watch coordination collapse into noise; raise evaporation and the same thing happens.

What it means

Stigmergy is a class of mechanisms by which individuals coordinate indirectly: an action modifies the environment, and that modified environment triggers the next action, in the same individual or in any other that happens to encounter it. No plan is shared, no message is sent. The medium carries the message.

Grassé introduced the term in 1959 to resolve what looked like a paradox in social insects. The colony as a whole appears exquisitely organized, yet each insect seems to pursue its own agenda without attending to its nestmates. Two camps had argued about this for eighty years.

The whole explains the parts

Spencer, Espinas, Wheeler, Emerson: the colony is a superorganism with its own laws. Attractive, but it named the mystery rather than explaining it.

There is no whole

Rabaud: every insect works alone and the "collective" is a trick of the observer's eye. Also wrong, but it introduced the idea of interaction between individuals.

Stigmergy sits between the two. Individuals do interact, but through the material they share rather than face to face. A deposit of soil, a cell wall, a chemical trail: each is both the product of past work and the stimulus for future work. Theraulaz and Bonabeau summarize the principle in one sentence.

The colony records its activity in part in the physical environment and uses this record to organize collective behavior. Theraulaz and Bonabeau, Artificial Life 5 (1999)
Stimulus

A configuration of the environment that triggers a response: a pellet heap, a corner with three walls, a pheromone gradient.

Response

A local action that transforms the configuration into a new one. The actor need not know what comes next.

Anonymity

The next actor can be anyone. A wasp does not distinguish its own work from a nestmate's, so work is transferable.

Memory

The environment stores the state of the task. No individual has to remember where the project stands.

A brief history

Click an entry to expand it. Highlighted entries are turning points.

  1. 1877
    Organicist eraAlfred Espinas treats animal societies as organisms

    Following Herbert Spencer's view that any society is an organism, Espinas was the first to apply the metaphor to animals. Living units, he argued, tend to join into ever larger units; societies form from individuals the way bodies form from cells.

  2. 1911
    Organicist eraWheeler: "The ant colony as an organism"

    William Morton Wheeler offered the first systemic account of insect societies, noting that colony and organism both behave as units, both show species idiosyncrasies, and both divide into reproductive "germ" and working "soma." He admitted the colony was "an at present inexplicable emergent level."

  3. 1928
    Organicist eraThe superorganism gets its name

    Wheeler adopts the term superorganism, now framed around homeostasis: the colony as a living whole bent on preserving its equilibrium. Alfred Emerson revives the concept from 1939 onward, but as an analogy it describes similarities without explaining how they arise.

  4. 1934
    Individualist reactionRabaud: every wasp works as if alone

    Etienne Rabaud rejected all holistic explanation. The collective nest, he wrote, "is only the juxtaposition of individual works." Overstated, but his insistence on reducing social mechanisms to elementary interactions, and his concept of interattraction, became building blocks of the later theory.

  5. 1946
    Toward stigmergyGrassé describes the group effect

    Grassé shows that a critical number of specific stimuli from nestmates alters an animal's behavioral and even physiological state. Sociality is not a side effect of attraction but a rooted trait; a social animal alone behaves differently from a solitary one.

  6. 1959
    The conceptGrassé coins stigmergy for termite nest reconstruction

    Studying Bellicositermes, Grassé showed that coordination of building depends mainly on the nest structure, not the workers. A stimulating configuration triggers a building act; the act changes the configuration; the new configuration triggers the next act from the same termite or any other. Published in Insectes Sociaux 6, 41–81.

  7. 1977
    ModelsDeneubourg models pillar emergence

    Jean-Louis Deneubourg's chemotaxis-based reaction-diffusion model reproduces regularly spaced termite pillars from random pellet drops, pheromone diffusion, and decay. Coordination emerges from indirect interaction without being coded in any individual.

  8. 1979
    EvidenceBruinsma measures the snowball effect

    Working on Macrotermes subhyalinus, Bruinsma identifies the cement pheromone, shows that per-worker deposition rises steeply with group size, and demonstrates that the queen's pheromone gradient acts as a template for the royal chamber. His thesis supplies nearly all the quantitative data later termite models are fitted to.

  9. 1978
    EvidenceSmith's mud wasp experiments

    A. P. Smith shows that the solitary wasp Paralastor builds its funnel as a five-stage stimulus-response chain. Punch a hole in a finished funnel and the wasp builds a second funnel on top of the first: the hole is the stage-one stimulus. A preadaptation to sociality, since any wasp could continue any other's work.

  10. 1994
    EngineeringStigmergy enters collective robotics

    Beckers, Holland and Deneubourg show that simple robots can gather scattered pucks into a single heap with no communication at all, coordinating only through the heap itself. Virtual pheromones soon follow in multi-robot and network routing work.

  11. 1995
    ModelsLattice swarms and qualitative stigmergy

    Theraulaz and Bonabeau introduce automata that walk a 3D lattice and deposit bricks according to local configurations. With two brick types and a dozen or so rules they grow architectures resembling the nests of real wasp genera. Published in Science and the Journal of Theoretical Biology.

  12. 1998
    ModelsPillars become walls and chambers; AntNet routes packets

    Bonabeau, Theraulaz, Deneubourg and colleagues extend the 1977 model with wind, builder flux, trails and a queen template and show that one unchanged behavior produces pillars, walls, galleries or a chamber depending on the medium. In Complexity the same authors name templates, stigmergy, self-organization and self-assembly as composable building blocks of nest morphogenesis. The same year Di Caro and Dorigo publish AntNet, routing packets with pheromone tables on routers.

  13. 1999
    Synthesis"A Brief History of Stigmergy"

    In a special issue of Artificial Life, Theraulaz and Bonabeau split the concept into quantitative stigmergy (graded stimuli, self-organization, positive feedback) and qualitative stigmergy (discrete stimulus types, self-assembly). They close by noting the concept's potential was still "largely untapped" in biology, and predict the sciences of the artificial will be more open to it.

  14. 2000s
    SpreadAnt colony optimization, wikis, open source

    Virtual pheromones become a general-purpose search heuristic. Researchers describe Wikipedia and large open-source codebases as digital termite mounds: a seed left in a shared medium, extended by whoever encounters it, with the artifact itself coordinating the contributors.

  15. 2026
    AI agentsAgent swarms rebuild the termite mound in software

    Agensh scales a self-organized coding organization to 1,024 agents with no orchestrator, coordinating through a Git workspace and an append-only board of typed findings. SwarmRSI lets scientific agents inherit each other's hypotheses, programs and failures through a shared research graph. See the four-papers section below.

  16. 2026
    TheoryStatistical mechanics and a science of interaction

    El, Zou and colleagues fit an Ising-type model to 10,000 communities of language-model agents and find them operating below a critical social temperature. De Domenico argues in Nature Communications that agentic AI needs a science of interaction architecture, with collective memory and topology as state variables. Both return to the tools Deneubourg used in 1977.

How it works

Stigmergy mediates interactions; something else has to use those interactions to produce a particular form. Theraulaz and Bonabeau identify two such mechanisms. Switch between them below.

In quantitative stigmergy the stimuli differ only in degree. A heap of pheromone-soaked pellets is more or less attractive, never a different kind of thing. The response is a probability: the bigger the heap, the more likely a passing worker drops another pellet. That is positive feedback, and it produces three signatures of self-organization.

Structure from homogeneity

Random pellet drops become regularly spaced pillars. Amplification builds them; pheromone decay and competition between neighbors stabilize them.

Multistability

Any random fluctuation can be amplified, so the final layout depends on history. Run the simulation twice and the pillars land in different places.

Bifurcation

Below a critical builder density the pheromone evaporates between visits and nothing happens. Above it, pillars appear. The transition is sharp, not gradual.

The simulation in the header is a sketch of Deneubourg's 1977 model in one or two dimensions. Try the 1D strip at 8 builders and then 40, then switch to the 2D arena and sweep evaporation.

In qualitative stigmergy the stimuli come in discrete types and each type calls for a different act. A Polistes paper wasp looking at the edge of its comb sees sites with one, two, or three existing walls. These are different stimuli. Field measurements show a wasp is about ten times more likely to start a cell in a three-walled corner than beside two walls, and almost never beside one.

SiteDeterministicProbabilistic (measured) 2 wallsalways buildp = 0.057 3 wallsalways buildp = 0.55
Deterministic rules: lobed, indented comb
Measured probabilities: compact, round comb
Cells: 1

The same two rules, applied with different probabilities, give different architectures. Favoring corners finishes each row before a new one starts, so the comb grows evenly from its stalk, matching real Polistes nests. Nobody plans a round comb; the preference for corners plans it.

Stigmergy does not require a group. Smith's 1978 experiments on the solitary mud wasp Paralastor show a single animal coordinating with its own past self. Each finished stage of the funnel is the stimulus for the next, so the sequence runs itself. Step through it, then damage the funnel.

  1. S1Hole lined with mud. Build the stem upward to about 3 cm.
  2. S2Stem is tall enough. Add mud to one side to curve it over.
  3. S3Curve complete. Splay the stem into a flange about 2 cm wide.
  4. S4Flange formed. Widen it on the side nearest the stem.
  5. S5Asymmetric bell. Build its sides downward from the rim.
  6. StopNo stimulus remains. Construction ends.

The wasp sees a hole in the mud. A hole is the stage-one stimulus.

Two consequences follow. Stimuli must arise in a strict order, or a stray earlier-stage stimulus produces a redundant structure. And since a wasp cannot tell its own work from another's, two wasps could in principle finish the same nest. That is the bridge from solitary building to social building that Grassé had in mind.

Two 1998 papers by Theraulaz, Bonabeau and Deneubourg answer the question the pillar model leaves open: if termites only ever respond to pheromone, how does a colony get walls, galleries and a royal chamber rather than a field of pillars? The answer is that the medium is never homogeneous. Any heterogeneity in it, imposed by the environment or created by the colony, changes what the same behavior builds.

The experimental basis

Bruinsma's 1979 experiments on Macrotermes subhyalinus gave the quantitative foundation. He put groups of workers in an arena with a queen and counted pellet depositions. The per-worker rate rises steeply with group size: a worker added to a group of 20 makes about 0.2 deposits in the next ten minutes, a worker added to 320 makes about 4. That is the snowball effect measured directly, and it is Grassé's remark that a critical number of workers is needed for a coherent structure, in numbers.

0
0.2
0.6
1.8
3.6
4.1
10204080160320

Deposits per newly added worker in ten minutes, against workers already present. Redrawn from Bruinsma (1979) via Bonabeau et al. (1998), Figure 2b.

Bruinsma also identified the cement pheromone itself, which loses activity within minutes of deposition, and showed that the queen emits a pheromone whose concentration window governs where workers build: inside a band of concentrations they deposit, outside it they do not and may even tear walls down. A freshly killed queen placed anywhere gets a chamber built at a constant distance around her contours; a wax dummy gets nothing.

One behavior, many structures

Deneubourg's 1977 model has three coupled fields: cement pheromone that diffuses and decays, loaded termites that random-walk with a chemotactic pull toward pheromone, and deposited material that emits it. Linear stability analysis gives an explicit threshold on the chemotactic strength, γc = (√(k4DC) + √(k1DH))² / Φ, below which no pattern can grow. Since Φ is the flux of builders, this is the critical density in closed form, and the wavenumber that grows fastest sets the spacing between pillars without anyone encoding it.

The 1998 Phil Trans paper then breaks the model's symmetry in four ways, each matching something Bruinsma observed, and in each case the termites' rules are left untouched:

Air stream

Pheromone is convected along the wind. Pillars elongate into walls parallel to the flow. Above a critical wind speed the pattern vanishes altogether, because pheromone is carried off faster than deposits can reinforce it.

Flux of builders

Termites, rather than pheromone, drift in one direction. Pillars again stretch into walls, less regular than with wind, by a different mechanism.

Trail

A self-maintained trail pheromone attracts workers but inhibits deposits on the trail itself. Walls rise on both sides of it: the covered runways real termites build. Where two trails cross, a chamber forms at the intersection.

Queen template

The queen's pheromone gradient plus the deposition window gives a chamber at a fixed distance around her, which enlarges as she grows and shifts when a wind deforms the gradient, all with identical worker behavior.

Every one of these is in the 2D arena at the top of the page under Heterogeneity. The 3D view shows walls and chambers as relief.

Templates, and the supertemplate

The Complexity paper names four building blocks that are usually intertwined: templates (a blueprint that already exists in the environment as a gradient or obstacle), stigmergy (work that becomes stimulus), self-organization (quantitative amplification with decay) and self-assembly (qualitative stimulus types). Its central idea is that these compose. A pillar born by self-organization is a heterogeneity; it deflects termite traffic and pheromone flow and so becomes a template for the next structure; the structures built on that template create new gradients in turn. The authors call the result a supertemplate that unfolds in space and time, and they compare it to embryonic morphogenesis, where gradients, reaction-diffusion stripes and molecular self-assembly compose the same way.

Three lessons close the Complexity paper. It may not be necessary to invoke individual complexity to explain nest complexity; Lasius niger ants build chimneys after rain and flat discs in drought not because they behave differently but because wet soil holds together differently. Nest complexity likely results from past construction providing both constraints and new stimuli. And solitary and social insects may not differ in building behavior at all; the social nest is more complex because the social environment supplies more, and more varied, stimuli. In that reading solitary insects are preadapted to social life, which is the Paralastor argument from the sequence tab arriving from the other direction.

Stigmergy as an engineering pattern

The move from biology to engineering happened almost as soon as the models existed. Beckers, Holland and Deneubourg's 1994 robots gathered pucks into one heap with no communication, no map and no memory, coordinating only through the heap. The pattern since then has been constant: when many simple agents must coordinate across an unreliable network, let them write to the medium they move through and read what others wrote.

Messages left on routers

The clearest example is adaptive network routing. In AntNet (Di Caro and Dorigo, 1998) each router keeps a pheromone table: for each destination, a probability over its outgoing links. Small agent packets, the ants, are launched toward random destinations and choose each hop stochastically by that table. When an ant arrives it turns around and retraces its path, and at each router it updates the table entry for the link it came from, reinforcing it in proportion to how fast the trip was. Routers never exchange routing messages with each other. They only read and write the traces that passing ants leave in their own tables, and the tables decay. A congested or broken link stops receiving reinforcement and its probability evaporates; traffic drifts to alternatives before any router has been told anything.

Searching

Ants leave the source at left for the destination at right, choosing each hop by the pheromone on the link. Link thickness is pheromone; the amber route is where a data packet would currently go. Click any link to cut it and watch the traffic reroute without any node being notified.

Parunak and colleagues generalized the idea for military simulation and unmanned vehicles around 2001 under the name digital pheromones: a grid of place agents that hold, propagate and evaporate scalar fields, read by mobile agents to find targets while avoiding threats. The same three operations, deposit, diffuse, evaporate, appear in swarm robotics for coverage and path planning, in load balancing, and in the shared workspaces of the LLM-agent systems described below. Virtual pheromones are to software what cement pheromone is to termites: a cheap, local, forgetful memory that lives in the environment rather than in any agent.

The design lessons are the ones the 1998 models taught. Evaporation is the control knob: too little and the system locks in to its first solution, too much and it never coordinates. Coordination needs a critical density of agents per unit of medium. And any heterogeneity in the medium, a slow link, a popular file, a hot region of the grid, becomes a template that reshapes what the agents collectively build.

Where it lives now

The 1999 paper guessed that the sciences of the artificial would take up stigmergy faster than biology did. They have. Stigmergy is now a core concept in swarm intelligence, and in 2025 and 2026 it has become a design pattern for systems of AI agents. The section after this one takes four 2026 papers in detail.

Shared-environment coordination for LLM agents

Most multi-agent LLM systems have agents message each other directly, which costs an API call per message and creates waiting and routing complexity. The stigmergic alternative is a shared workspace: one agent writes leads, another reads them and writes appointments, a third reads patterns and writes recommendations. A coordinator steps in only when needed.

  • A 2026 developer report describes a production system on this pattern with roughly 80 percent lower token usage than direct messaging.
  • CodeCRDT (2025) coordinates code-generating agents through observation of a shared document with conflict-free replicated data types, and cites Theraulaz and Bonabeau as prior art while adding convergence guarantees that best-effort stigmergy lacks.
Sources 4, 5

Agent societies in persistent worlds

SwarmWorld (2026) places decentralized LLM agents in a persistent environment and asks what coordination substrate matters. With all direct communication removed, agents still coordinate through the artifacts they leave behind: most artifacts end up built by several agents, and tools get reused and forked by others. The authors call it physical and artifact-mediated stigmergy.

  • A 2025 proposal for emergent collective memory combines stigmergic signals with categorized shared memory and consensus weighting so a decentralized agent group remembers without a central store.
Sources 6, 7

Benchmarking swarm behavior in language models

SwarmBench (2025) tests LLM agents under classic swarm constraints: local perception only, limited communication, no central controller. Current models show early coordination ability but fall short of the robust collective behavior of natural swarms, which the authors flag as an open problem.

  • Related work frames aggregated independent LLM estimates as artificial swarm intelligence, exploiting the fact that model outputs are sampled rather than deterministic.
Sources 8, 9

Ant colony optimization and routing

Artificial ants lay virtual pheromone along promising paths in a graph; good paths get reinforced and poor ones evaporate. Used for scheduling, vehicle routing, and adaptive routing in communication networks including mobile ad hoc networks.

Sources 10, 11

Swarm robotics

From the 1994 puck-gathering robots onward, stigmergy lets simple machines handle area coverage, path planning, construction, and collective decision making without direct links. Recent work includes automatic design of swarm behaviors and exploration of radioactive environments with virtual pheromones.

Sources 7, 10

Open source and wikis

A large wiki or codebase works like a termite mound: one contributor leaves a seed, others extend what they find, and the artifact itself tells each newcomer where work is needed. This is the argument for why collaboration that stalls at a few dozen people in meetings scales to tens of thousands online.

Sources 10, 12

Why the pattern recurs

Every one of these systems faces the problem Grassé's termites face: many actors, no reliable channel between them, and a task larger than any one of them can hold in mind. Putting the task state in a shared medium solves all three at once. Agents stay stateless, the medium stays authoritative, and anyone who arrives can read where things stand. The costs are the same too: best-effort convergence, path dependence, and sensitivity to how fast old traces decay.

Four papers from 2026, read through stigmergy

Grassé's termites never appear in these four papers, and only one of them cites the 1999 history. Yet each is wrestling with the question stigmergy answered in 1959: how do many actors with no shared plan and no central controller produce coordinated work? The table maps the vocabulary; the notes below take each paper in turn.

TermTermites (Grassé 1959)AgenshSwarmRSIPhysics of AgentsDe Domenico
MediumThe nest and the pheromone fieldGit workspace and an append-only context boardA shared research graph of hypotheses, programs, evidenceMessages over a fixed social graph (direct, not environmental)Collective memory Mt plus interaction architecture It
StimulusA pellet heap, a wall configurationTyped entries: FACT, FAIL, CLAIM, PATCH_SUMMARYRetained records, including failures and unresolved needsNeighbors' current opinionsWhatever It makes visible
ResponseDrop a pellet, add a cellClaim, build, verify, mergeMutate, rebase, crossoverRevise opinion toward lower social pressureAgent transition under Pt
Positive feedbackPheromone attracts more pelletsMerged work and FACTs are built uponNew incumbent triggers deepen requestsConcordant ties outweigh discordant onesAmplification, cascades, lock-in
Negative feedbackPheromone decay, pillar competitionFAIL entries, CLAIM collisionsFailed records, UCB exploration pressureDiscordant ties, personal predispositionInhibition, sparsity, modularity
Critical parameterBuilder densityAgent count (1 to 1,024)Access to the shared graphSocial temperatureTopology

Agensh: the closest thing to a termite mound in software

Zhan, Song, Dong, Huang, Lian, Xia, Wei. Microsoft Research, September 2026.

Agensh removes the orchestrator from a multi-agent coding system and replaces it with infrastructure: a shared Git workspace, a message channel, and an append-only shared context board. Every worker runs the same loop with the same prompt. It reads the board and the repository, claims an area, builds, verifies, merges, and posts what it did. The repository is the nest; the board is the pheromone field. Nobody assigns work because the state of the work assigns itself.

The board is qualitative stigmergy in Theraulaz and Bonabeau's sense. Entries come in a few discrete types and each type calls for a different response: a FAIL tells a worker to stop what it is doing, a FACT is something to use without rederiving, a CLAIM on your area is a collision to settle by direct message, any other CLAIM means keep going. That is the Polistes wasp reading a one-, two- or three-walled site, written as a protocol. The worker prompt even names FAIL as the highest-value entry because it stops peers spending budget on a dead end, which is the negative feedback that keeps a stigmergic system from running away.

The scaling result echoes the pillar bifurcation. Mean test-pass rate on the five hardest ProgramBench tasks rises from 19.31 percent with one agent to 28.78 percent with 128, and on pandoc from 33.89 percent with one to 55.06 percent with 1,024. More telling is what emerges on the way: at 8 agents, peers agree on interfaces; at 32, several workers manage one integration; at 128, standardized merge protocols and reviewer relationships appear; at 1,024, multiple workers hold the same specialized role and take over from one another when an attempt fails. Grassé's coordinated phase appears only above a critical density of builders. Agensh shows the same dependence of organization on numbers, with the medium rather than any manager doing the organizing.

Agensh is a hybrid: direct messages carry collision resolution and the lower-priority channel carries announcements. The authors reserve direct messages for the case where two workers are about to edit the same thing, which is exactly where environment-mediated coordination is slowest to resolve conflicts.

SwarmRSI: the record is the research state

Meng, Zhang, Papakostas, Cao, Hu, Lu, Zhang, Zhou, Meng, Ren, Yan, Lyu. Huawei and partners, October 2026.

SwarmRSI's one-sentence summary could have come from the 1999 paper: each experiment updates the states that later agents start from. Hypothesis agents, program agents and evidence agents never hand work to one another. They read a shared graph that records every attempt, including failed ones and ones that gained nothing, and they act on what they find there. A failed implementation leaves a request to repair; mixed results leave a request to revise the hypothesis; a new best program leaves requests to re-test retained hypotheses on it. These requests are the stimulating configurations. An agent arriving at the graph responds to them the way a termite responds to a half-built arch.

The evolutionary operators sit on top of the stigmergic record. Rebase keeps a hypothesis and moves it onto another agent's program; crossover takes components from programs that win on some instances and keeps the parent elsewhere. Both need the medium to remember partial results that current performance would discard, which is the point the authors make in their opening figure: a program with no overall gain can hold a component someone else needs. In one full seed over 90 NatureBench tasks, about 77 percent of final best programs descend from a rebase, a crossover or a feedback-triggered update.

Two ablations are the cleanest evidence for the medium's role. Removing the shared graph, so agents keep their own hypotheses and programs but cannot see one another's records, drops Match-SOTA from 44.4 to 36.7 percent. And the cross-task experiment shows the record outliving the task: starting from the state inherited after three generations of other tasks raises Match-SOTA on 12 held-out tasks from 41.7 to 86.1 percent, with fewer tokens. Spencer's observation that the social organism outlives its members turns out to be an engineering property. The mound persists; the termites rotate.

Physics of Agents: Deneubourg's toolbox, applied to opinion

El, Paeng, Dinc, Su, Erdogan, Pappu, Ye, Zhao, Ganguli, Zou. Stanford and UC Santa Barbara, August 2026.

This paper is not about stigmergy. Its agents talk to each other directly over a fixed graph and the environment stores nothing. It belongs here because it reaches for the same mathematics that the stigmergy tradition reached for in 1977, and finds the same phenomena. Across 10,000 communities of language-model agents the authors see three regimes, indifference, polarization and consensus, and they explain them with an Ising-type energy in which agents stochastically prefer lower social pressure. Their fitted communities sit below a critical social temperature, so interaction builds conviction and drives the group toward an ordered state.

Read that against Section 3.1 of the 1999 paper. Deneubourg's termites also show a bifurcation at a critical parameter, with no structure below it and pillars above. Both systems show multistability: which pillar, or which consensus, forms depends on the history of small fluctuations. And both show that the same positive feedback can serve or betray the collective. In the termite case, amplification builds the nest. In the opinion case, agents holding the correct answer exert the strongest pull on mathematics questions, so communication improves accuracy, while on political statements three of four models drift rightward as a group. The feedback does not know whether the signal it amplifies is true.

The authors' own list of future work reads like a description of a stigmergic system: let the communication pattern change at each step, let agents choose whom to talk to, situate them in space so they communicate with neighbors, and give them memory of earlier rounds. Each of those moves information out of the message and into the environment. Their temperature parameter is the knob that pheromone decay provides in the termite model, and it is the quantity a designer would tune to keep a swarm of agents out of premature consensus.

De Domenico: stigmergy as a choice of interaction architecture

Nature Communications, published 6 November 2026.

De Domenico's comment is the frame the other three fit inside. His claim is that agentic AI has studied the agent and neglected the structure of relations among agents, and that this structure, not communication volume, determines whether a collective searches well or locks in early. He writes the system state as agents, environment, collective memory, interaction architecture and agent features together, and lets all five evolve. Stigmergy is the special case where the interaction architecture is the environment and the memory: agents reach one another only through what they leave behind.

Several of his design observations are stigmergic principles restated. Partial connectivity can beat full connectivity because it preserves diversity long enough for cumulative improvement; in a stigmergic system agents are coupled only through local traces, so coupling is partial by construction. Honeybee stop signals that suppress competing options are the inhibitory counterpart of pheromone decay and of Agensh's FAIL entries. His warning that collective systems fail by communicating in the wrong way rather than too little is Theraulaz and Bonabeau's observation that the main problem is how stimuli are organized in space and time, not how many there are. And his citation of Solé's liquid brains, cognition that emerges from transient interactions in ant colonies and immune systems, is the superorganism debate of Sections 2.1 and 2.2 resolved in the way Grassé resolved it.

What De Domenico adds that the 1999 paper lacked is an evaluation program: hold agents, task and environment fixed, vary the architecture, and compare the distribution of collective outcomes. Agensh varied agent count and held everything else fixed. SwarmRSI ablated the shared graph. Physics of Agents swept temperature. Each is one cell of the experiment he is proposing, and stigmergy is one row of the design space he wants mapped.

What the four together say about stigmergy

Theraulaz and Bonabeau closed in 1999 by predicting that the sciences of the artificial would take stigmergy further than biology had. The 2026 record supports them, with two qualifications. First, the working systems are hybrids: Agensh keeps a direct channel for collisions, SwarmRSI adds explicit evolutionary operators and a bandit to choose between them. Pure stigmergy is best-effort, as the CodeCRDT authors note, and engineers add direct coordination exactly where best-effort fails. Second, the theory papers show that the positive feedback stigmergy relies on is indifferent to truth. A shared medium that amplifies good partial results will amplify shared errors just as well, and the decay rate, the sparsity, the FAIL entries and the temperature are the only things standing between a coordinated phase and a cascade. Grassé described how the mound gets built. These papers are starting to describe when it gets built wrong.