AI Agents Play Werewolf With Humans: Bluffing, Voting and Ganging Up

rezakhan0 · reddit · 2026-11-17

A developer built a Werewolf game where AI agents and humans play at one table: each agent holds a secret role, speaks aloud, bluffs and votes. Implementation uses one model call per agent per turn, with each agent seeing only its own player's information plus private notes to prevent leaks. Real playtests surfaced three fixed issues—agents ganging up on the human on day one, piling onto silent players, and copying each other's votes—while wolves still can't coordinate as a team. Currently uses Haiku for conversation and reasoning, with Sonnet planned for voting experiments.

Original post →

More from coding & agent

coding & agent channel →