Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati

Why watch

Erina Karati presents Project Paradox, a game agent framework built at Supercell's AI Innovation Lab with memory, emotions, and trust scores. The video explains how the team solved long-horizon failures (agents conflating rumors with facts) by implementing an autoresearch loop that runs controlled scenarios and scores agents on reach, source retention, uncertainty preservation, replanning, and privacy. This is valuable for AI engineers building autonomous agents that must maintain accurate information over extended interactions.