Supercell's Project Paradox: long-horizon agents need experiments, not just prompts
AI Engineer · youtube · 2026-09-27
In an AI Engineer video, former Microsoft/Supercell engineer Erina Karati presents Project Paradox, built with Arunachalam Manikandan at Supercell's AI Innovation Lab — a framework giving game agents memory, emotions, and trust scores.
The problem: it worked in short scenes but broke down over long horizons. A rumor about a mango sale degrades as it spreads — agents forget who said it and turn "might" into fact.
The fix: an autoresearch loop
- Run controlled scenarios (spreading a fact, a rumor, a plan change) and collect traces.
- Score results on a balanced scorecard: reach, source retention, uncertainty preservation, replanning, and privacy.
- The loop may only edit a small, frozen policy surface, and changes are kept only if the scorecard improves.
Lessons beyond games
- Memory alone isn't enough — you must know where each fact came from.
- Rollback isn't optional.
- Long-horizon agents need experimental loops, not just prompt tweaks.
The video includes full demos (a picnic scenario, the fixed rumor) with timestamps.
More from coding & agent
- OpenAI selects 35 builders for first Codex Physical Builds hardware cohort — OpenAIDevs · 2026-09-27
- Claude Opus 5.5 Chains Blender, Image-Gen and three.js Into Browser-Running 3D — burny_tech · 2026-09-27
- Claude kills plan mode the same day Google announces launching it — dotey · 2026-09-27
- Gary Marcus amplifies warning: large teams using AI agents likely have unknown security incidents — GaryMarcus · 2026-09-27
- Viral speech: writing code by hand is no longer economically productive — AccBalanced · 2026-09-27
- He recreated the ChatGPT UI with Creatr, and his February mockup mostly held up — koltregaskes · 2026-09-27