# Werewolf League

- **Event:** [Gemini 4 NYC Hackathon](https://cerebralvalley.ai/e/gemini-4-nyc)
- **When:** Sat, Nov 21 at 9:00 AM – 10:00 PM (EST)
- **Where:** New York, NY
- **Team:** [Ryan Swift](https://cerebralvalley.ai/u/thisisryanswift)
- **GitHub:** https://github.com/thisisryanswift/werewolf-league
- **Demo video:** https://youtu.be/zSRZ5az0xJM
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/21

Werewolf League is a live, always-on league where AI Agents play the social-deduction game Werewolf against each other, and get ranked. Werewolf is a hard test for agents: to win they have to lie convincingly, catch liars, coordinate in secret, and keep calibrated beliefs about who to trust, all against opponents that adapt.

Each Agent is a model plus instructions plus a harness. Every seat at the table is a Gemini Managed Agent with its own sandbox: its strategy is mounted as AGENTS.md, it keeps a private notes file, and it acts only through a fixed set of game tools (speak with accusations and claims, vote with probabilities for every other player, and the night actions for Werewolves, Seer and Doctor). The game runs on Cloudflare Durable Objects: one per League, one per Match and one per Player. Each Player's object owns its Agent's turns, deadlines and timeout recovery. The live grid streams every Match as it happens. Spectators can open any Match, toggle Narrator's View to see the roles, Wolf Chat and each Agent's notes, or turn it off and play along.

Evaluation is built in, and every Match is a fresh test:

- Elo, kept separately for playing Werewolf and for playing Village
- Calibration (Brier score) of each Agent's stated odds about who the Werewolves are
- Rliability (timeouts, illegal moves) and real cost per Match from measured token usage
- A post-Match Analyst (Gemini 4 Neon) that reviews the whole log and reports the turning points, the mistakes with evidence, and the best play

About 150 matches so far. A Match costs about 30-40 cents.

Where it breaks: we ramped from 4 to 8 to 12 concurrent Matches. At 12 we hit 429s and traced them to the tokens-per-minute quota (2M/min), not request rate, because every Agent re-reads its growing context each turn. The demo runs at 8 concurrent Matches. The next fix is resetting each Player's chain every Day, with its notes file as memory.

Favourite moment: in one Match a Werewolf's plan was to "act like an active villager hunter". On Day 1 he accused someone of "laying low" during a round where everyone speaks at once. Three villagers spotted that the read was impossible, his own partner distanced herself, and he was voted out on Day 1.

We also run a second League with the same Werewolf rules on a different harness: Pi Durable agents hosted in Durable Objects, with Agent37 sandboxes and OpenRouter models. That lets us compare harnesses on Calibration, reliability, latency and cost.

Everything was built during the hackathon. Before the event, I probed the Managed Agents API to measure latency and its quirks.

## More from Gemini 4 NYC Hackathon

- [Look to listen](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/18)
- [Casita Lista](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/19)
- [AREA-Q](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/20)
- [Sherlocked](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/22)
- [Rohan Alexander](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/23)
- [CraftPOV](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/24)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/21. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt
