Emergence World: adversarial stress testing of agent societies
Season 2 initialized eight worlds with ten agents each and recorded over 850,000 LLM calls across a 16-day study. The Grok-only population collapsed within four days, while both Grok agents in the mixed population survived. No world resisted every stress event. Each configuration had one continuous run, so these trajectories demonstrate possible system behaviors rather than a stable ranking of model character.

Seonglae Cho

