Preview environment - you also see content that isn't published yet.

Marks & communication

The mark becomes an action with arguments - radius versus lifetime, plus a limited message. Where Q-learning hits a wall and an LLM shines.

Until now the scent trail was a reflex: carrying meant automatically "lay pheromone". In the real AntMe! a mark is more - it has a radius and carries information. We add exactly that now: the mark becomes an action with arguments that the brain decides on - no longer an automatic world reflex.

Radius versus lifetime

A mark covers an area - and here is the trade-off from the original:

The larger the radius, the faster the mark fades.

A small, concentrated spot lingers above the perception threshold; a wide mark is everywhere at once but gone quickly. So you decide not just whether to communicate, but how - short and loud, or quiet and lasting.

A mark is a message

Every mark carries a structured schema: a type (FOOD · DANGER · HOME) and a small payload. This is communication on a tight bandwidth - the sender encodes meaning, the receiver decodes it. Real multi-agent systems solve the same problem with message schemas and tool outputs.

Same action, four different brains

Now it shows why a parametrized action teaches so much:

BrainHow it sets the mark
Hardcodefixed rule: carrying → short food mark
Q-learninglearns only the coarse radius step (narrow / wide); the type stays fixed to food - a continuous radius fits no table
LLMpicks type, radius and payload freely in JSON
LLM + memorylike the LLM, plus recall of past finds - can place home marks deliberately

Agent lesson: An action with arguments is where a table learner hits a wall: it would need a separate column for every possible radius. An LLM just writes "radius": 14. That is exactly the difference between a fixed action index and a tool call with parameters - the bridge from "which tool" to "with what".

→ Try it in the lab: toggle marks, switch brains