Intel / video

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

Source: youtube.com
Author: AI Engineer
Date: Published Sep 26, 2026

Context

An can remember a message while losing the information needed to judge it. Project Paradox, a simulated village of AI characters, encountered that problem when a rumor passed between agents: its source disappeared, or a possibility became an asserted fact. The project talk describes testing those failures through controlled conversations and changes of plan.

The improvement loop scores whether information spreads while preserving attribution and uncertainty, alongside the agents' ability to revise plans and respect privacy. It can change only a restricted set of behavior rules, and keeps changes only when the scorecard improves. This makes the quality of a remembered claim, including who supplied it and how certain it was, part of the optimization target.

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.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source www.youtube.com