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

In a game village full of AI agents, a rumor about a sale on mangoes goes wrong. Pass it along a few times and the agents forget who said it, or start treating "might" as fact. Erina Karati, a former engineer at Microsoft and Supercell, built Project Paradox with Arunachalam Manikandan at Supercell's AI Innovation Lab. It's a framework for game agents with their own memory, emotions and trust scores. It worked well in short scenes but broke down over long ones. Their fix was to wrap the village in an autoresearch loop. They run controlled scenarios like spreading a public fact, a rumor or a change of plan, collect traces, and score the results on a balanced scorecard: reach, source retention, uncertainty preservation, replanning and privacy. The loop can only change a small, frozen policy surface, and changes are kept only if the scorecard improves. Karati's lessons apply well beyond games: memory alone isn't enough, you need to know where each fact came from, rollback isn't optional, and long-horizon agents need experiments, not just prompts. Speaker info: Erina Karati X/Twitter: @erinakarati (https://x.com/erinakarati) Website: https://www.erinakarati.dev/ Arunachalam Manikandan X/Twitter: @Arunachala64250 (https://x.com/Arunachala64250) Related links: Supercell: https://supercell.com Timestamps: 0:00 Intro 0:43 Project Paradox 1:18 What the agents can do 2:53 A stateful architecture 3:33 Trust scores and memory importance 4:28 Demo: a picnic 5:08 Where long-horizon behavior breaks 6:43 Bringing in autoresearch 8:13 Autoresearch outside the village 9:23 The loop 10:43 Controlled scenarios 12:33 The mango rumor, fixed 13:03 A balanced scorecard 14:33 Keep the editable surface small 15:43 Example policy changes 16:33 Being careful with claims 17:03 Memory is not enough 17:48 Rollback is not optional 18:23 Beyond games 19:33 A recipe for long-horizon agents 20:13 Closing