Agentopia: Long-Term Life Simulation and Learning in Agent Societies

A team of researchers has introduced Agentopia, a framework that simulates 100 autonomous agents over 10 years to study emergent social behaviors and train language models to better replicate human social intelligence. The work, posted to arXiv on June 5, 2026, addresses a limitation in prior agent society simulations, which typically operate at the scale of days, restricting the depth of social interactions and long-term growth [1][2]. Agentopia instead runs agents across three diverse worlds for a full decade of simulated time, during which agents autonomously pursue personal growth, develop social relationships, and fulfill needs and goals [2][3]. The researchers define a "life reward" metric designed to mirror human well-being and use it to train the underlying large language model through rejection sampling, selecting high-advantage agent trajectories for optimization [2][3]. Experiments show that agents exhibit rich emergent behaviors, including personal growth, relationship formation, and life choices [3]. The life reward training improved agent well-being in simulation, measured through broader social recognition, higher subjective fulfillment, and better economic gains [3]. The trained model also generalized to downstream role-playing benchmarks, yielding a 15.6% overall improvement on the CoSER Test, with the largest gains in anthropomorphism (+23.7%) and character fidelity (+16.4%) [3]. The Agentopia framework enters a growing field of long-horizon multi-agent simulations. Other recent systems include EvoSpark, which integrates narrative control and cognitive evolution to sustain coherent story worlds, and AIvilization v0, a public-facing societal platform where users steer autonomous agents via long-term goals [4][5]. TaskWeave, another framework, models organizations as role-structured agent populations and uses a hierarchical simulation-memory mechanism to maintain planning and execution states over year-long enterprise simulations [7]. These systems share a common challenge: maintaining stable agent identity and coherent behavior over extended time horizons, a problem Agentopia tackles through its decade-scale design and life-reward optimization loop [2][5]. arXiv, the open-access repository where the paper appears, hosts more than two million e-prints and receives roughly 24,000 submissions per month as of late 2024, making it a central venue for pre-publication research in computer science and related fields [11]. The Agentopia code has been made available on GitHub [3].