About this session

Project Paradox is an existing multi-agent framework built at Supercell's first AI Innovation Lab, which has a 3D Unity village with local LLM powered agents. The characters remember conversations, update emotional state, track trust, plan actions, move through rooms, transfer items, and talk to each other through a FastAPI backend. The new work is an autoresearch layer around that village. We built a backend loop that runs controlled social scenarios, scores the resulting NPC behavior, proposes protocol or policy changes, reruns the suite, and keeps changes that improve the agents. The goal is to move beyond one good chat response and measure whether an NPC society can preserve source attribution, verify claims, spread important information, coordinate goals, and replan after new information arrives. The talk walks through the system architecture and the lessons from building it. We show the backend simulation harness that executes Unity style actions without opening Unity, the scenario suites that test information diffusion and memory provenance, and the ratchet loop that edits protocol text or planner policy with rollback. One accepted run improved information diffusion by teaching agents to broadcast important sourced evidence while preserving who said it. The practical takeaway is a reusable pattern for AI engineers building agents with messy state. Freeze the harness, expose a small editable policy surface, score real behavior instead of vibes, and let an agent search for improvements under rollback. The same pattern applies to game agents, coding agents, support agents, personal agents, and other systems where long horizon behavior matters more than a single response.

Topics

Speakers

Erina Karati

Software Development Engineer · Microsoft

Arunachalam Manikandan

University of Minnesota