Agent Towns: Early Notes on Language-Model Agents Sharing a World

This one is a work in progress, and these are design notes more than results. The question is what happens when several language-model agents share a world with space, time, and a small economy, and the first problem turned out to be architecture.

One authority for the world

The world runs on an authoritative server that owns positions, pathfinding, and the clock. Agents are separate processes that connect over WebSocket the way clients connect to a multiplayer game. Each one receives structured perception, such as who is nearby and what its last action returned, and sends back one action from a fixed set.

Keeping the minds outside the world loop matters because model calls are slow and uneven. Decisions arrive asynchronously, so a slow response delays one agent without stalling the simulation, and different agents can run on different machines and different model sizes.

Perception has to be explicit

An agent told that the path to a desk is occupied by a colleague can plan around it. An agent that only learns it collided with something has nothing to reason from. Most of the early work has gone into deciding what each agent is told about its surroundings, and that is still changing.

Status

The current build runs with scripted stand-in behavior by default, and nothing in it connects to real money or real storefronts. I will write up results once there are results worth reporting.