No agent calls another directly. World, souls, and players publish and subscribe on one bus — the same topology runs single-process today and distributed tomorrow.
Owns canonical state · validates every consequence · closes each turn.
One per character — private memory, secrets, agendas. They perceive, decide, and act concurrently.
One per player — each renders its own narrative projection of the same world.
The paper shipped with 時間陣 (Jikanjin), a playable multi-agent world demonstrated at the EC2026 session in Kyoto. The same runtime powers WorldLines today: every character its own agent, the world agent narrating, an agent map tracking who is where.
The formal model — Parameterized-Action POMDP, canonical JSON state, and the PDVA transition kernel.
The society layer — world, soul, and player agents as peers on a pub/sub bus, with the playable demo 時間陣.
The growth paper — play expands the canonical world through validated transitions; it grows without drifting. Preprint after review.
Overview of all three: worldlines.gg/research →
@inproceedings{huang2026livingworlds,
title = {Multi-Agent LLM Simulation for Living Game Worlds: Coordinating
World, Soul, and Player Agents over a Pub/Sub Message Bus},
author = {Huang, Yuhang and Li, Chenmiao and Fang, Chaowei},
booktitle = {IPSJ Entertainment Computing Symposium (EC2026)},
address = {Kyoto, Japan},
year = {2026}
}
WorldLines — the multi-agent engine for living worlds. File-backed, event-sourced, playable today.