Agentopia: Teaching AI to Live Like Humans in a Decade-Long Simulated Society

This article introduces the Agentopia framework published on arXiv in June 2026 by the team of Xintao Wang et al. The system enables 100 AI agents to live autonomously for a decade in a simulated society. Through long-term social interaction and a training mechanism based on life rewards, the agents display rich emergent social behaviors, and the underlying large language model achieves a 15.6% performance improvement in role-playing benchmark tests.

Human intelligence stems from long-term social life learning, but previous multi-agent simulations only lasted a few days or weeks, lacking depth and growth trajectories, and thus unable to answer the core question of whether AI can evolve from long-term social experience. Agentopia aims to address this issue through a decade-long simulation.

The framework includes three key parts: 1. Agent Architecture: Based on LLM, with cognitive architecture, personal profiles, and a needs system (simulating human hierarchy of needs); 2. Social Environment: Contains various places, with environmental states changing over time; 3. Time Progression: Event-driven mechanism focusing on key decision points.

During the decade-long simulation, naturally occurring behaviors were observed: 1. Stable social relationship networks (friendships, romantic relationships, families, with dynamic evolution); 2. Career development trajectories (career choice, promotion, career change, entrepreneurship); 3. Group cultural phenomena (trend propagation, intergenerational value transmission).

The 'life reward' is defined to quantify the agent's life satisfaction (need fulfillment, goal achievement, social relationships, etc.), referencing positive psychology. A rejection sampling method is used to fine-tune the LLM based on reward feedback, judging the quality of decisions to optimize the model.

Experimental results: 1. The average life reward of agents in the simulation increased steadily; 2. A 15.6% improvement in downstream role-playing benchmarks; 3. Ablation experiments show that both long-term simulation and life reward training are critical, and the capabilities are transferable.

Three major challenges and their solutions: 1. Computational Efficiency: Event-driven approach + importance sampling to control costs; 2. Agent Consistency: Personality encoding to enhance behavioral coherence; 3. Evaluation Difficulty: Multi-dimensional quantitative + qualitative analysis, combined with human evaluation.

Significance: Opens up the direction of long-term multi-agent simulation, demonstrates the possibility of AI learning from experience, and provides an experimental platform for AI safety. Future directions: Expand simulation scale, introduce complex environments, apply to fields such as economics and urban planning.