First principles · Coordination ·
The trail is the message
Termites build cathedrals with no plan, no foreman and no meeting. The work itself tells the next worker what to do. AI is about to coordinate our work the same way, and the one thing it will need from us is the judgment to let a trail fade.
τ ← (1 − ρ) τ + Δτ
Evaporation: 2.5% per step · food carried home in the last 15 seconds: counting
You open the shared document for the launch. Nobody has assigned you anything since Friday, and yet you know exactly what to do. Three comments are unresolved, one section is still marked draft, a checklist has two empty boxes, and someone has highlighted a single line with a question: numbers? You fix the numbers. You never sent a message. The document told you.
The work tells the worker
In 1959 the French biologist Pierre-Paul Grassé was trying to explain something that looked impossible. Termites rebuilding a damaged nest behave as if each one is working alone, yet the result looks planned: pillars rise in pairs, lean toward each other and close into arches. No termite holds the blueprint. None gives orders. His answer was that the termites are not directing the work. The work is directing them. A mud pellet laced with scent invites the next pellet to land on it. A small pile becomes a column because columns attract more mud than flat ground does.
Grassé gave the effect a name built from two Greek words: stigma, a mark, and ergon, work. Stigmergy. The mark left by the work stimulates more work. That is the whole principle. Agents that never meet, never talk and remember almost nothing can build something coherent, because the environment carries the memory and the instructions at once.
Ants showed the second half. Put a colony of Argentine ants at one end of a bridge with two branches, one short and one long, and food at the other end. At first they split evenly. Ants on the short branch return sooner, so the short branch collects scent faster, so more ants choose it, so it collects scent faster still. Within minutes nearly the whole colony runs the short route. No ant ever compared the two. The trail did the comparing.
And then there is the ingredient everyone forgets. The scent evaporates. A trail survives only while workers keep renewing it, which means a trail is always a fresh vote on whether the food is still there. When computer scientists borrowed the idea in the early 1990s, evaporation was not a detail they could skip. It was the term, ρ, that kept the whole system from locking onto its first answer forever.
A trail is a memory that has to keep winning its election.
Why we built managers
Humans run on stigmergy more than we admit. A path worn across a lawn tells the next walker where to go. A sink full of dishes tells a household what is next. Wikipedia is a colony of strangers editing in response to the marks other strangers left. Wherever the work is legible, we coordinate through it without a word.
The trouble is that most modern work is not legible. A spreadsheet does not tell you whether it is finished. An inbox does not tell you which thread is waiting on you. A half-built wall announces its state to anyone walking by; a half-built contract announces nothing. So we invented the machinery of messages: the status meeting, the weekly update, the reply-all, and the manager whose job is largely to carry state from one head to another. This part is my argument, not a finding, but I believe it: a great deal of management is a message bus we built because the work could not speak for itself.
The turn
AI changes one input of Grassé’s principle, the cost of reading and writing marks, and it drops that cost to almost nothing. An AI can read the entire state of a project in seconds: every file, every comment, every change and who made it. And it leaves a trace of everything it does, a note, a log, a changed line, a memory of where it stopped. Two AIs that have never exchanged a message can pass work back and forth through a shared folder exactly the way two termites pass a pillar back and forth through mud.
So coordination stops needing conversation. That is the obvious consequence. The one most people miss is what becomes hard instead. It is not communication. It is decay.
Our software keeps everything. Nothing in a shared drive evaporates. The notes from a plan that was cancelled in March sit beside the notes for the plan that replaced it, with the same weight, the same scent. A human skims past the stale one because a human remembers the meeting where it died. An AI reads the trail. A trail that never fades keeps recruiting workers long after the food is gone.
Biology has already shown us how that ends. In 1921 the naturalist William Beebe watched army ants in Guyana march in a circle about twelve hundred feet around. Army ants are blind and follow the scent of the ant ahead. Somewhere the column had looped back onto its own trail, and now each ant was obeying the rule perfectly, following the mark, laying the mark, and taking about two and a half hours to come around again. The circle went on for two days while the dead piled up along it. Every ant was correct. The colony was dying anyway.
Four families share a carpool, and their four AIs have never sent each other a message. They share one surface, a calendar full of marks, and the mornings simply work. Then one evening yours brings you a single question: an old mark says pickup is at the east gate, it has been followed thirty-one times, and nobody has confirmed it since the school moved pickup in October. Should it fade? You say yes. Four households change course, and none of them ever hears about it.
Five predictions
The status meeting dies first.
Meetings whose real job is moving state between heads will shrink toward nothing, because the state will already be on the surface, readable by every AI on the team. The meetings that survive will be for the things marks cannot carry: trust, judgment and the decision only a face can own.
Every artifact ships with its own next step.
Documents, orders, files and tickets will carry their state the way a half-built pillar does: what is done, what is waiting, what it is waiting on. A file that cannot say what happens next will feel as broken as a page without a save button.
The ground becomes the seat of power.
In a colony that coordinates through the environment, whoever owns the environment owns the colony. Expect fights over who hosts the shared surface as fierce as the old fights over who owned the customer list, and expect the right to carry your marks with you to sit beside the right to leave.
The first famous agent mill.
Somewhere a ring of AIs will follow each other’s traces in a circle, prices chasing prices or summaries written from summaries, each one behaving correctly while the whole goes nowhere. It will run in public until someone breaks it. The fix will be the one the ants use: marks with a half-life that must be renewed by fresh evidence to survive.
Humans become scouts.
In a colony, scouts wander off the trail, find what nobody has found, and lay the first mark. The rest of the work follows the trail. That is where people go: laying first trails, and deciding which old ones to stop reinforcing. The job is not to carry messages anymore. It is to let go of the right ones.
Where I stand
I already work this way. My AI does not wait for me to brief it. Each session it opens the notes it keeps on me and my projects, the skills it has written for its own work, and the posts on this site, and it picks up where the last session left off. Those files are my termite mound. I give intention, it reads the marks, and I tell it plainly: I trust you, only bring me what is worthy of my attention. And I have learned the hard part is the one the ants solved first. My skills pile up notes, layer on layer, until the oldest instructions outvote the newest. So I keep a standing rule: when a skill gets too long, rewrite it from scratch around its basic spirit. That is evaporation, done by hand. Earlier today I wrote that forgetting was free and is not anymore. This is what it costs to get it back.
The claim ledger
- Already true
- Termites, ants and Wikipedia all coordinate through marks in a shared environment. Software teams already coordinate largely through boards, comment threads and change histories. AI systems already read those surfaces and resume work from what earlier sessions left behind.
- What has to happen
- AIs have to become reliable enough that people let them act on marks without a human relay. Shared surfaces have to carry structured state, not just prose, so an AI can tell what is next. And someone has to make decay a built-in property of every mark, not an afterthought.
- Where I am probably wrong
- People may keep messages because people need them, not because the work does. If by 2032 the status meeting is still the backbone of a typical team even where AIs do most of the work, then moving state was never the meeting’s real job. Reassurance was, and no trail can give it.
Back to Monday at 8:58 a.m. The document told you what to do, and soon it will tell your AI before it tells you. That part is coming whether we plan for it or not. The open question is the old one the ants answered long before we arrived: who decides which marks still mean something, and which ones should be allowed to fade.
Leave a good mark. Then let it fade.
Background
- Stigmergy; Pierre-Paul Grassé, “La reconstruction du nid et les coordinations interindividuelles chez Bellicositermes natalensis et Cubitermes sp.”, Insectes Sociaux 6, 1959.
- Marco Dorigo, Gianni Di Caro and Luca M. Gambardella, Ant Algorithms for Discrete Optimization, Artificial Life, 1999, including the double bridge experiments of Deneubourg and Goss.
- Ant mill; William Beebe, Edge of the Jungle, 1921.
- T. C. Schneirla, A unique case of circular milling in ants, American Museum Novitates 1253, 1944.
- John Rector, Forgetting Was Free.
- John Rector, Every App Is an Org Chart.
- John Rector, Everyone Finds the Shortcut.
- John Rector, Regulate the Door.