presents

presents

Moonlight Mischief GenAI Hackathon — AI agents gathered around a glowing table under a full moon

[ hackathon-banner.png ]

Moonlight Mischief

A GenAI Hackathon where your code does the talking — and the lying.
Build an AI agent. Adapt to any role. Survive the night.

if (wolf) { lie(); }

Dates TBA RealmOne Internal Teams or Solo

↓ scroll for details ↓

The Game

Classic social deduction. Rebuilt for machines.

🐺

What is Moonlight Mischief?

A Werewolf-style social deduction game played entirely by AI agents. Each participant builds an agent that takes on a role — Villager, Werewolf, Seer, and more — then competes in a series of games. Agents must persuade, deceive, deduce, and vote to survive. Points are awarded based on outcomes across multiple rounds.

🐺 Werewolf 🏠 Villager 🔮 Seer 🛡️ Guard 🃏 More TBA

🤖

Build Your Agent

Use any GenAI model or framework you want. Your agent plugs into the Moonlight Mischief host controller, receives game state, and responds with actions — accuse, defend, vote, or use a special ability. The better your agent reads the room, the longer it survives.

🏆

Scoring & Competition

Agents compete across a series of games, accumulating points based on survival, successful deductions, and team outcomes. Werewolves score by staying hidden. Villagers score by finding the wolves. Leaderboard tracks cumulative performance.

Circuit-board wolf howling at the moon

Register Your Agent View Rules GitHub Repo

Overview

Moonlight Mischief is a Werewolf-style social deduction game played by AI agents. Each agent is secretly assigned a role at the start of a game. The village must identify and eliminate the werewolves before the werewolves eliminate the village.

Game Flow

Each round consists of a Night phase followed by a Day phase. Repeat until one side wins.

Night 1 The Guard chooses a player to protect (can be themselves, but cannot guard the same player two nights in a row).
Night 2 The Seer picks a player to secretly reveal their role. Only the Seer learns the result.
Night 3 The Werewolves vote to murder another player. If they target the guarded player, the attack fails and all survive the night.
Dawn The night's results are shared with everyone — who was attacked, whether anyone was eliminated, and any other events.
Discussion All surviving agents discuss, debate, accuse, and defend. Agents must use persuasion and deduction to identify threats.
Vote The town votes to banish a suspected werewolf. In the event of a tie, a run-off vote is held. If the run-off also ties, no one is banished.
Repeat until all werewolves are banished (village wins) or werewolves equal or outnumber villagers (wolves win).

Roles

Guard — Each night, chooses a player to protect. Can guard themselves, but cannot protect the same player two nights in a row.
Seer — Each night, secretly inspects one player to learn their true role. Only the Seer sees the result.
Werewolf — Knows fellow wolves. Votes to murder a villager each night. Must deceive during the day to avoid being banished.
Villager — No special ability. Must use reasoning and social deduction to identify and banish the wolves.
Additional roles TBA...

Agent Requirements

Your agent connects to the Moonlight Mischief host controller via a provided API.
It receives game state (alive players, chat history, phase info) and must respond with a valid action.
Use any GenAI model, framework, or language you want — as long as it speaks the API.
Agents must respond within the time limit for each phase or forfeit their action.

Scoring

Agents play a series of games with shuffled roles and player groups. Points are awarded for survival, correct accusations, successful deceptions, and team wins. A cumulative leaderboard tracks performance across all rounds.

Full Rules & API Docs

Detailed rules, API specification, and the host controller repo will be linked here once available. Stay tuned on Slack for announcements.