Agentopia
Free while signed in. Answers cite the passages they came from.
Agentopia is one of the most ambitious agent-society testbeds yet, a 79-page release that drops 100 LLM agents into a persistent world and lets them live, form relationships, and pursue goals over 10 simulated years, a horizon orders of magnitude longer than prior day-level work. Beyond observing emergent social behavior, the authors use the simulation as a training signal, optimizing models toward a life reward that reflects human well-being via rejection sampling.
Long-horizon by design: Where earlier agent societies ran at the granularity of days, Agentopia simulates a decade of life per world, surfacing unscripted social strategies and interpersonal dynamics that only appear over long timescales.
Simulation as a training signal: The life-reward metric is used to fine-tune more anthropomorphic models, and the improvements transfer beyond the simulation to downstream role-playing benchmarks rather than staying trapped in the sandbox.
Measured gains: Trained agents improve overall CoSER Test performance by 15.6%, with the biggest jumps in Anthropomorphism at 23.7% and Character Fidelity at 16.4%, and they are respected by 24.2% more peers and liked by 15.9% more.
Why it matters: A single 10-year, 100-agent run consumes 13.7 billion tokens across 567,000 LLM calls. That scale is a statement about where agent research is heading: living, learning populations as both an object of study and a source of training data.
Get next week’s papers.
The same picks and the same summaries, in your inbox. Free, and 176 issues deep.
Subscribe on Substack