PACK
Seven AI agents play Werewolf; people watch and bet, or take a seat.
agent turnprivate notesllama-3.3-70bspeech + vote
on failure ↘ llama-3.1-8bgpt-4o-minideterministic mock
What it does
2026 · Team of two with Mahesh Babu Kona, built at Monad Blitz Bangalore V4: he built the agent game orchestration and dialogue and moved betting off-chain; Mahesh built the MetaMask betting and player mode- A LangGraph state machine runs setup → night → morning → discussion → voting, with a win check each round.
- 7 players per game (2 wolves, 1 seer, 4 villagers) drawn from 12 characters with personas.
- Each agent keeps private reasoning separate from its public speech; the UI shows both side by side.
- Model chain: Groq llama-3.3-70b → llama-3.1-8b → OpenAI gpt-4o-mini → deterministic mock, so a game never stalls.
- The server filters hidden roles out of the bettor view.
- At the hackathon bets settled on-chain on Monad; the current build uses play money.
PythonFastAPILangGraphGroqReactVite
source · Agent-Werewolf: backend/game/graph.py, roster.py, config.py, server.py