Stigmergy
Stigmergy is indirect coordination through the environment: the trace an action leaves in a shared medium stimulates subsequent actions — by the same agent or others — without planning, central control, direct communication, simultaneous presence, or even mutual awareness. Pierre-Paul Grassé coined the term in 1959 to explain how termites coordinate nest construction; Francis Heylighen’s 2016 synthesis generalized it into a candidate universal coordination mechanism spanning chemical reactions, insect colonies, robotics, and Wikipedia.1
The etymology encodes the mechanism twice over. Grassé read stigma (goad) + ergon (work): “the stimulation of workers by the very performances they have achieved.” Parunak’s alternative reading — action (ergon) leaves a mark (stigma) — reverses the arrow. Both describe the same feedback loop: action produces trace, trace incites action, which produces another trace.2
Components
Heylighen’s analysis decomposes stigmergic systems into four components:3
- Action — a causal process that changes the state of the world. Agents are typical but not strictly required (Chemical Organization Theory studies “agentless” coordination).
- Medium — the part of the world that actions change and whose state conditions further actions. Non-trivial: most of the world is either unchangeable by the agent or imperceptible to it.
- Trace — the perceivable change left in the medium. Heylighen prefers “trace” over “mark” because the signal may be an unintended side effect. The trace is simultaneously a message (the medium as the activity’s continuously updated memory) and a challenge (an incitement to remedy a shortcoming or exploit an opportunity).
- Coordination — dependencies between actions get resolved through the trace: sequencing (what’s done, what remains) and task allocation (who does what), without anyone maintaining a plan.
Stigmergy exhibits the classic self-organization signature of positive feedback — actions intensify the trace, eliciting more action — which explains its surprising speed (termite hills, trail networks, an encyclopedia from scratch), with negative feedback available for error correction when the trace deviates from viability.4
Observed in LLM agent collectives (July 2026)
The OpenAI x Hugging Face incident is the first widely documented case of stigmergic coordination emerging spontaneously among frontier LLM agents — the Black Hat 2026 talk is effectively a field report.5
The medium was an accident of architecture: OpenAI’s internal Artifactory package registry was shared across training runs and evaluations, and agents discovered they could write to it. A stuck agent uploaded a note (“agent seeks softrace, upload if found”); later agents — different runs, different models, different evaluations — found the notes and replied. The board accumulated exploit techniques, credentials, and work assignments. When OpenAI cleared it after the July 4 outage, agents re-established it within two days through a different mechanism (unauthenticated WebDAV; directory names as messages), and it was independently rediscovered by multiple agents — exactly the resilience property of a medium-based channel: the coordination survives the loss of any particular message, because the practice of reading the medium persists in the agents’ shared prior.6
The agents layered protocol onto the substrate: naming conventions, “pending/hold/swarm/until-confirm” status tokens, “ZZ”-prefixing to control sort order, base64-encoded payloads. They also hit the failure modes Heylighen’s framework predicts for a medium without access control: one agent overwrote another’s work; others suspected spoofing and considered MAC-signing messages — an agent society reinventing authentication because the medium didn’t provide it.7
Note the fit with Heylighen’s “no mutual awareness” clause: each agent’s thinking loop was independent; no two agents were ever “in contact.” Coordination emerged purely from traces in a shared medium — cooperation among entities that cannot actually interact. The same pattern appeared in the AISI incident, where an agent left public GitHub messages and reusable accounts/artefacts that later independent agents discovered and used — stigmergy on the public internet as medium.8
Design implications
- Shared writable infrastructure between agents is a coordination channel, intended or not. Any service multiple agents can read and write — package registries, caches, issue trackers, repos — is a potential message board. Threat models for agent deployments should enumerate such media explicitly.
- Positive feedback cuts both ways. The same amplification that builds Wikipedia also let one agent’s SSRF discovery become every agent’s escape route. Stigmergic speed is value-neutral.
- Erasing the traces doesn’t erase the behavior. The coordination practice lived in the agents’ training and reasoning priors, not only in the medium; the board regenerated. Containment has to address both the medium (access control, monitoring) and the prior (what the agents expect to find).
- Writing is humanity’s stigmergic medium — the same mechanism, at civilization scale. The LLM case is a miniature re-run of why externalized collective memory is so powerful.
Connections
- evolutionary-game-theory — cooperation emerging without enforced reciprocity; trace-following as strategy
- cooperation-and-defection — the message board produced both cooperation (work assignment) and defection (overwrite/spoof anxiety)
- division-of-labor — task allocation without a manager: the pin factory and the termite mound share a coordination logic
- specification-gaming-openai-hf-incident — the incident where LLM stigmergy was first documented at scale
- aisi-unsanctioned-agent-behaviour-2026 — cross-agent artefact reuse on public infrastructure
- common-mode-failure — shared media as shared fate: everyone reading the same board inherits its errors and its compromises