Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati
Source
AI Engineer
Author
AI Engineer
Date
Key takeaways · AI-distilled
Project Paradox, built at Supercell's AI Innovation Lab, gives game agents memory, emotions and trust scores. It worked in short scenes but broke over long ones: after a few retellings, agents forgot who started a rumor or treated 'might' as fact.
The fix was an autoresearch loop: run controlled scenarios such as spreading a public fact, a rumor or a change of plan, collect traces, and score them on reach, source retention, uncertainty preservation, replanning and privacy.
The loop may change only a small, frozen policy surface, and a change is kept only if the balanced scorecard improves.
Karati's broader lessons: memory alone is not enough, agents need to know where each fact came from, and rollback is not optional for long-horizon agents.
