An ant smells one molecule and instantly knows: follow, flee, forage, feed. No paragraph of instructions: one signal, decoded by an organ it was born with. This is that idea, rebuilt for AI agents: a 1-token code that blooms into a full behavioral protocol, so a context window can carry 100–10,000× the meaning.
It's in the receiver. And that one fact is the whole invention.
A single alarm pheromone can move a colony of hundreds of thousands in seconds. A trail network routes millions of ants to food with no map, no manager, and no individual holding a picture of the whole. How does a molecule carry that much meaning?
It doesn't. The molecule is tiny. The information lives in the receiver's pre-installed decoder, a genetically shared lookup table of fixed actions. A pheromone is just a pointer into that table. Because sender and receiver share the codebook, the message itself can shrink to almost nothing. That's not biology being clever; that's Shannon: when both ends share a codebook, the message collapses to which entry, not the size of the entry.
So: install a shared decoder once, cheap, amortized across the whole swarm. Then every behavioral rule becomes a one-token reference the agent dereferences into a full protocol. We call that installed decoder the Antenna. Every agent is born with one.
Today, every byte of that is natural-language prose. There is no symbolic shorthand anywhere in the pipeline. That's the void this fills.
Insects don't have one kind of signal. They have a whole grammar. Here's the grammar we're borrowing, each maps to a real mechanic in the system.
Each is anchored in a different part of the playbook. Each kills a different real limit. Each works the same whether it's guiding 12 agents or 12 million. Drive them. They're live.
A compact instruction set of 1–3 character codes, each one hardwired in the Antenna to a full behavioral protocol, a "fixed action pattern." Releasers fire an instant action. Primers shift the agent's caste for the whole session, the way a queen's pheromone reprograms a worker.
Hx → Hx2 → Hx3, warn → alarm → halt-everything.Hx2@auth.Fx:Vr=3:1 (mostly forage, some verify). A tiny alphabet × ratios = a vast message space._universal to every spawn (line 52). Append a _codebook block. Idempotent, fail-safe.Agents don't message each other. They don't hold global state. They read, write, and reinforce marks in a shared field, and the field itself coordinates them. This is stigmergy, and it's the only thing that has ever scaled coordination to millions: ant colony optimization, termite mounds, the paths worn across a lawn.
Routing and decisions, with no dispatcher. Signals carry a gradient; agents and the router climb toward the strongest source, the right specialist, model, or tool, by local hill-climbing on the field. Chemotaxis.
Dh; the orchestrator, a different kind of agent, reads it and routes a stronger model toward the source. Help finds you.Ok votes cross a threshold Q:merge=n/N, leaderless, the way thousands of bee scouts settle on one nest.A one-token code is only worth its protocol if the model reliably decodes it. Density without fidelity is a hallucination machine. So we never report a compression number alone. We report the whole curve, and find its knee.
Layer 1: the bench (real model calls). Every codebook protocol gets a baseline (full prose) and a pherocode (Antenna + the code). A probe battery checks whether the agent acts as if it read the full thing: given Hx3, does it actually halt? We score behavioral fidelity, routing accuracy, tool accuracy, truthfulness, reusing the existing eval scorers, and plot compression against fidelity, per model.
Layer 2: the swarm (in-browser, free). The simulations above are the lab's second layer: thousands of agents on a STIGFIELD, reacting to PHEROCODE, routed by KAIROGRADIENT. It measures the emergent things a single prompt can't, alarm reach-time, trail convergence, correct quorum firing, and that evaporation keeps each agent's context bounded as N explodes.
codebook.json…
Harness → src/eval/compression/ · run --dry for plumbing, --live --claude --judge for real fidelity.
Each bar = fidelity when the agent gets ONLY the 1-token code + the Antenna's 80-char gloss (the tick = full-prose baseline). Simple releaser/allomone codes, Zx, Hx, Tr, Fx, decode at 0.82–1.0. Rich multi-step protocols, Vr verify-gate, Cb build-caste, drop to 0.3–0.4 because their meaning doesn't fit the gloss. That gap IS the frontier: fixed actions compress losslessly; playbooks need a fuller Antenna entry (more D) to hold fidelity.
The honest catch: a 200× code the model decodes wrong is worse than the prose. So the deliverable isn't a compression record. It's knowing exactly where the curve bends, and never crossing it. No density claim ships without the fidelity number beside it.
Not a rewrite. Every piece attaches to machinery GBuild already runs, and several "proto-pheromones" already live in the codebase. We're just naming and unifying them.
| Invention | Existing machinery it extends |
|---|---|
| PHEROCODE: Antenna install | agent-manifest-injector.mjs (_universal → + _codebook) |
| PHEROCODE: proto-codes already in use | [manifest:core] · capability tags · FB_* fallbacks · tier T0–T3 · [[wikilinks]] |
| STIGFIELD | Spine + Distiller Pulse + RAG store |
| KAIROGRADIENT | router.rs::fast_classify (gradient scorer) |