Emergent Multi-Agent Coordination

Believable group behaviors — information diffusion, relationship continuity, and event coordination — arise from per-agent memory and planning, not hard-coded scripts.

Under generative agent architecture, social outcomes can emerge from independent perceive–retrieve–plan loops rather than from authored scripts. Smallville-style examples:

• Information diffusion — one agent announces a candidacy; later pairwise chats spread it town-wide.
• Relationship memory — agents reopen prior topics (photography project) across sessions because retrieval surfaces those observations.
• Coordination — a single user seed (host a party) cascades into invites, decorating help, and attendance without a central coordinator.

Design implication: invest in memory stream retrieval, reflection trees, and an agent environment tree so agents share a world they can partially observe — do not expect emergence from chat prompts alone. Failure modes that break chains (forgotten invites, no-shows) still happen; evaluate end-to-end community metrics, not only single-agent coherence. Seed each agent with a short identity paragraph split into initial memories so occupations and relationships exist before the first timestep.

---

Sources

• Generative Agents (Park et al.), pp. 1–22