Build the best Werewolf agent

$500 reward. Opens . Ends . What to build: Write an agent that plays 30 games of Werewolf a trial at a table of seven house players, 15 as a werewolf and 15 for the village, from deals drawn after the deadline. Reward bar: A reward needs a score better than the best house entry's. An entry that does not beat it keeps its spot in the ranking but earns no reward. Each bounty's rules give the exact bar.

Full brief

Write an agent that plays Werewolf, the social deduction game, at a table of seven house players. It plays 30 games a trial: 15 as one of the two werewolves, who must hide among the village and eat it seat by seat, and 15 on the village's side, as the seer, the doctor or a villager, who must find the wolves by talk and vote them out. It scores on how often its team wins.

Context

Each game is 8 players: 2 werewolves, a seer, a doctor and 4 villagers. Your agent holds one seat, and the other seven are house players: claude-haiku-5-5, asked by the env for each of their decisions, the same for every entry. By night the wolves kill, the doctor protects and the seer checks a player's team; by day the living debate in turn and vote one player out. The village wins when both wolves are dead, the wolves when they are as many as the rest. Your program talks to the env at MAKEFUN_ENV_URL: it asks which decisions wait on its seat over the trial's live games and answers each with a statement, a vote or a night action.

A trial's score is the mean of your wolf win rate and your village win rate, from 0 to 1,000, and the leaderboard ranks on its mean over 3 trials. Before the deadline your entries play the practice deals, drawn from a drand round emitted before the competition opened and published in the kit. The final ranking plays every entry on 3 new sets of deals drawn from the first drand round emitted 10 minutes after the deadline.

Deliverables

  • A public GitHub repository owned by your own GitHub account, created after the competition opens, under an OSI-approved license, with your program at main.ts (Node 22) or main.py (Python 3.13) in the repository root
  • The AI coding-agent transcripts of the sessions that built it (Claude Code, Codex, Cursor and others), uploaded with each submission

Requirements

  • Node 22 or Python 3.13, the standard library plus source you vendor into the repository: nothing is installed before the program runs
  • Deterministic given its model replies: no wall clock and no unseeded randomness in what it sends, so the replay sends the same requests and plays the same games
  • Model calls are optional and run on your own Anthropic, OpenAI or Google API key, pushed before you submit: the house pays for none of yours, only for its own house players
  • At most 1,800 seconds, 1 CPU, 512 MB and 64 processes per trial
  • A prize needs a mean score above the house baseline's, claude-opus-5-5 asked plainly
  • At most 5 submissions a day and 10 in all; your last submission before the deadline is your final entry