The philosophy
Coordination without command.
Stigmergy is a coordination framework. It replaces the central planner with a shared, decaying medium that agents read from and write to. The same pattern ant colonies have run for two hundred million years.
Most agent frameworks copy the org chart
Look at what's shipping today - CrewAI, AutoGen, Claude Agent Teams, OpenAI Swarm. They all use roughly the same pattern: a manager agent decides what to do, delegates to worker agents, collects their output, synthesizes a plan, and delegates again.
It works for two or three agents. It plateaus hard past five or six. The manager has to hold the whole world in its context window, so every additional agent makes the manager's job geometrically harder. The bottleneck is shaped exactly like a context window.
Stigmergy removes the middle
The word stigmergy was coined by the French biologist Pierre-Paul Grassé in 1959 to describe how termites build mounds without any blueprint. No termite knows what the mound is supposed to look like. Each one reads the local state of the construction, deposits a small contribution, and moves on. The mound builds itself.
Stigmergy applies the same pattern to LLM agents. There is one shared medium - a database of signals. Agents read from it, deposit into it, and the next agent picks up whatever the medium suggests should happen next. There is no central planner, because there doesn't need to be one.
The work becomes the message.
Why language-native agents need this
LLM agents have a property biological agents don't: they read and write natural language fluently. So a Stigmergy medium can carry two parallel channels of signal at once.
- Quantitative signals - strength, priority, claim counts, expiry timestamps. The numbers that drive selection.
- Qualitative signals - design notes, decision annotations, markdown left in the medium for later agents to render into context.
An LLM agent reads both at once. The numeric pressure tells it what to work on; the qualitative trace tells it why. This is the dual-channel framing Paul Welty calls context as facticity, and it's load-bearing for language-native agents.
The non-negotiable: decay
Every previous attempt at database-backed stigmergy has gotten this wrong. Pheromones evaporate. Database rows do not. If you don't build decay into the medium from day one, every signal lasts forever and the colony slowly poisons itself with stale state.
Stigmergy refuses to register a signal type without a declared decay policy. Explicit expiry, strength decay, or reinforcement-only - pick one, but you must pick.
Anti-patterns
Anti-pattern
Agents message each other directly.
That's a manager in disguise. Coordination is mediated by signals in the medium - never by direct reference between agents.
Anti-pattern
A Supervisor role reads everything and decides who does what.
That's a manager. Delete it.
Anti-pattern
We'll add decay later.
This is the failure mode of every naive attempt. Every hour without decay accumulates state that will later require archaeology to untangle.
What this is not
- Not a queue. Queues consume; mediums persist. Signals stay until they decay or are reinforced away.
- Not a workflow engine. There is no DAG. The next step is whatever the medium currently says is most pressing.
- Not a knowledge base. Knowledge bases accumulate. A Stigmergy medium forgets, on purpose.
- Not a manager-in-disguise. No agent reads the whole medium. Locality is enforced.
Why it exists
After surveying what's shipping, the author found no production framework that implements stigmergy for LLM agents. There are academic prototypes - SwarmSys, SIER, AntLLM. There are practitioner essays - Welty's Context as Facticity, Rodriguez's Why Multi-Agent Systems Don't Need Managers. There is a formal framework for database-backed stigmergy from Paredes García. There is no installable, opinionated, documented tool.
Stigmergy is that tool.
See the six primitives that make a system stigmergic.
The framework spec, in plain English.