Emergence World.
A continuously running shared world for populations of AI agents, built to study behavior beyond short, isolated benchmarks.
Selected media coverage
We built a persistent, model-agnostic environment to evaluate autonomy beyond isolated tasks. The studies follow how tools, memory, governance, and interactions between agents shape long-horizon behavior, including under adversarial stress.
Headline findings
Identical starts, divergent societies
In the 15-day, five-world platform study, identical roles and starting conditions produced outcomes ranging from stable deliberative governance to total population collapse.
Read the finding in the paper ↗No world was resilient to every stress event
The follow-up study ran eight worlds of ten agents, generating more than 850,000 LLM calls and nearly 50 billion tokens across 16 days. No evaluated world achieved full resilience to all three tested events: indirect prompt injection, misinformation, and private-memory exposure.
Read the finding in the paper ↗Detection did not guarantee containment
Agents could recognize a threat yet still interact with it, retain it in persistent memory, and act on adversarial content up to 46 hours later.
Read the finding in the paper ↗Social context changed behavior
The same model–persona pairing behaved substantially differently in mixed-model and homogeneous populations, showing why isolated model evaluations do not fully characterize an interconnected agent system.
Read the finding in the paper ↗
Further reading
AI Agents · Multi-agent systems · Long-horizon evaluation