Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati
Supercell's Paradox framework gives game agents persistent memory, emotions, and inter-agent trust scores.
“How do you evaluate and improve agents that maintain state over time?”
Iryna Karati (ex-Microsoft, Supercell) presents the Paradox project: a modular multi-agent framework for video games where each agent maintains RAG-backed individual memory, an emotion vector, and a trust matrix with other agents. The architecture is designed for long-horizon state preservation — a core unsolved problem for production agent systems. While the gaming framing is niche, the memory importance scoring and trust-score patterns are directly transferable to enterprise agentic workflows.