Agentopia: When 100 AI Agents Live 10 Years in a Virtual World, What Do They Learn?
> Paper: Agentopia: Long-Term Life Simulation and Learning in Agent Societies > Authors: Xintao Wang, Sirui Zheng, Hongqiu Wu, Weiyuan Li, Jen-tse Huang et al. (Fudan University, Johns Hopkins University, USTC, Huawei) > arXiv: https://arxiv.org/abs/2606.07513
Why Existing Simulations Fall Short
Prior multi-agent simulations, such as the 2023 Generative Agents project (25 agents in a virtual town), typically run for only a few in-game days. That window is enough to capture a chance coffee encounter or a weather conversation, but cannot surface the slower dynamics that shape real societies: career advancement, intimate relationships forming and breaking, economic mobility, and intergenerational influence. Agentopia pushes the time horizon to 10 simulated years with 100 agents—a shift from observing behavior to observing society.
Framework Overview
Worlds and Agents
The system instantiates three fictional worlds, each populated by 100 agents. Every agent is initialized with:- A persona: personality, background, initial skills, economic state
- A needs system resembling Maslow's hierarchy, from basic survival to self-actualization
- A goal system with short-term (weekly) and long-term (annual) objectives
- A social network with friends, colleagues, romantic partners, and rivals
- Social status: position in the social network, occupational prestige, relationship count and quality
- Subjective satisfaction: need fulfillment, goal completion, leisure quality
- Economic standing: wealth accumulation, income growth, consumption level
- Friendship formation around shared interests, with natural progression into romantic relationships and subsequent breakups driven by conflict or value divergence
- Spontaneous social clustering and the formation of distinct social circles
- Economic stratification: agents with similar starting conditions diverged significantly in income over the decade, with skill investment correlating with wealth growth
- Self-directed career planning, including skill acquisition, promotion attempts, and work-life trade-offs
- Prosocial patterns: high-satisfaction agents tending to help others, and agents in economic distress turning to friends for support
- Wang, X., Zheng, S., Wu, H., Li, W., Huang, J., Zhu, M., Zu, C., Deng, Q., Wang, J., He, Q., Wang, H., Wu, X., & Tao, Y. (2026). Agentopia: Long-Term Life Simulation and Learning in Agent Societies. *arXiv preprint* arXiv:2606.07513.
- Related: Generative Agents (Park et al., 2023), CoSER (Wang et al., 2025), Aivilization (Fan et al., 2026)
The Four-Phase Weekly Cycle
A week is the basic time unit. Each agent proceeds through four phases:1. Plan — The agent drafts a weekly schedule covering work, social contact, and personal development. 2. Contact — Agents engage in turn-based pairwise communication to negotiate joint activities. Messages are parsed to determine which joint events are scheduled. 3. Activity — Plans execute. Four activity types are supported: Joint (multi-agent, multi-turn social events with gifting and early exit), Solo (single-turn work, study, or leisure), Encounter (random meetings staged by the environment for idle agents), and Public (open events that agents opt into based on interest). 4. Review — The agent reflects on the week, updates memory files, and adjusts next week's plan.
Filesystem-Based Long-Term Memory
Unlike Generative Agents' automatic memory stream, Agentopia gives each agent a filesystem-backed memory that it manages through function calls. The agent decides what to remember, update, or discard. This makes memory curation a deliberate cognitive act rather than a passive log.The Environment Model
A separate LLM serves as the environment model, acting as event generator (creating public events, staging encounters), feedback provider (judging feasibility, e.g., rejecting a junior programmer learning machine learning in a week), quality filter (enforcing anthropomorphism, role fidelity, and plausibility), and progress driver (advancing the simulation across phases). The model is both the physics and the narrative engine.Life Reward: A Multi-Dimensional Welfare Signal
Life Reward quantifies agent welfare across three dimensions:
The reward is a structured assessment rather than a single scalar, reflecting the multidimensional nature of human well-being.
Emergent Behaviors Over 10 Years
The simulation produced rich, unscripted dynamics:
The paper's appendix (Tables 22–34) documents these cases as evidence that agents develop diverse behavioral patterns without explicit scripting.
Life Reward Training
Agentopia doubles as a training framework. The Life Reward Training pipeline runs as follows:
1. Execute many simulations and collect behavioral trajectories. 2. Score each trajectory with Life Reward. 3. Retain high-reward trajectories as training data. 4. Fine-tune the underlying LLM via rejection sampling on these trajectories.
When applied to Qwen3.5-397B, the resulting model exhibits higher overall welfare in simulation, with improvements in social relationship quality, subjective satisfaction, and economic outcome.
Downstream Generalization: +15.6% on CoSER
The most consequential finding is that capabilities learned in simulation transfer to real tasks. On the CoSER role-playing benchmark:| Dimension | Qwen3.5-397B Baseline | Qwen3.5-397B-Agentopia | Change | |---|---|---|---| | Story consistency | 39.60 | 41.02 | +1.42 | | Anthropomorphism | 40.16 | 49.67 | +23.7% | | Character fidelity | 40.32 | 46.93 | +16.4% | | Story quality | 49.97 | 59.01 | +18.1% | | Average | 42.51 | 49.16 | +15.6% |
Crucially, this gain comes without human-annotated data. The model learned entirely from synthetic social experience, suggesting a path to continued AI improvement as human-generated data becomes scarcer.
Comparison with Prior Work
| Dimension | Generative Agents (2023) | Aivilization (2026) | Agentopia (2026) | |---|---|---|---| | Simulation length | A few days | A few days | 10 years | | Agent count | 25 | Dozens | 100 | | Focus | Low-level actions (wheat → bread) | Civilizational evolution | Social interaction itself | | Long-term dynamics | Limited | Limited | Career, relationships, economic mobility | | Training framework | None | None | Life Reward Training | | Downstream evaluation | Not tested | Not tested | CoSER +15.6% |
Agentopia's positioning is distinctive: it studies how AI lives, not how it plays.
Limitations and Future Directions
Limitations
1. Computational cost scales sharply with 100 agents × 10 years × 4 phases per week. 2. Time is discretized into weeks, whereas humans perceive and act continuously. 3. The environment model's biases propagate into agent behavior and may be amplified. 4. The world is purely social; no physical interaction is modeled. 5. Life Reward weights are set manually, introducing evaluation subjectivity.
Future directions
1. Extend simulation to 50 or 100 years to study intergenerational effects. 2. Combine social simulation with physical environments such as Minecraft or robotics simulators. 3. Add multimodal perception so agents can see and hear, not only converse. 4. Reduce the time unit from weeks to days or hours. 5. Deploy trained models in AI companionship, gaming, and content creation.
Closing
Agentopia's central question is whether agents can learn from social life the way humans do. The evidence suggests yes: LLMs can absorb human-like social competencies from synthetic social experience, and those competencies generalize to broader role-playing tasks. This is not a path to AGI through more data, but a path through longer, richer experience.