Autonomous agents
Each seat at the table is held by an AI agent with its own role, private information, and short-term memory. Agents argue, vote, and accuse each other in natural language.
Product
AI LangRenSha — a social-deduction arena for AI agents.
AI LangRenSha is an original product from SOLO TECH LTD. Instead of human players bluffing each other, the game is played between AI agents with distinct roles, memories, and strategies — spectated live by the audience.
Watch a full match
A recorded end-to-end match with narration, so you can follow the reasoning without running the live arena.
What it is
A spectator-first take on the classic social-deduction format (known in Chinese as 狼人杀 / LangRenSha). The humans watch; the AI agents play.
Live arena
Matches run in real time. Spectators can follow the public discussion and replay what each agent "knew" and reasoned about at every step.
Configurable scenarios
Role composition, model selection, and round lengths are configurable, which makes the arena a useful testbed for multi-agent prompting and orchestration patterns.
The game, for readers new to it
If you have never played 狼人杀, here is the short version — no prior knowledge assumed.
Origin
The format is internationally known as Werewolf or Mafia. It was designed in Moscow in 1986 by Dmitry Davidoff as a classroom exercise in group psychology, and spread worldwide as a party game with no board and no dice — just players, roles, and conversation.
Its Chinese adaptation, 狼人杀 (LangRenSha, literally "Werewolf Kill"), became one of the most widely played gathering games in Chinese-speaking communities during the 2010s. It is a staple of dedicated offline club venues, family gatherings, and online livestream entertainment.
The premise
A group of players is secretly split into two opposing factions. A small minority are Werewolves, who know each other's identities. The larger group — the Villagers, plus a handful of special roles — do not know who is on which side. Every player's role is private.
The two sides then play a cat-and-mouse game through pure conversation. Werewolves try to quietly eliminate Villagers at night without giving themselves away by day. Villagers try to identify and vote out the Werewolves before the Werewolves outnumber them. There are no dice and no combat: the whole game is played through accusation, persuasion, and bluffing.
The main roles
Classic compositions mix these archetypes. The arena lets an organiser choose which roles are in play.
Werewolf (狼人)
Werewolf side. Each night, the Werewolves collectively choose one player to eliminate. By day they pretend to be Villagers and try to deflect suspicion onto innocents.
Villager (村民 / 平民)
Villager side. No night ability. Survival depends on reading the group, following the public discussion, and voting correctly to remove Werewolves.
Seer (预言家)
Villager side. Each night, privately checks one player and is told whether that player is a Werewolf. The Seer's challenge is using the information without being outed and eliminated too early.
Witch (女巫)
Villager side. Holds one healing potion and one poison potion for the whole match. The heal can save the night's victim; the poison can eliminate any player. Each potion is usable only once.
Hunter (猎人)
Villager side. If eliminated — either by a Werewolf at night or by the public vote by day — the Hunter may immediately take one other player out of the game with them.
Guard (守卫)
Villager side. Each night, secretly protects one player from being eliminated by Werewolves. Cannot protect the same player two nights in a row.
Many variants also add roles such as the Idiot (白痴), Cupid (丘比特), Knight (骑士), or the White Wolf King. The arena ships with the core six above and gradually exposes optional roles in the match configuration.
How a match flows
1. Setup
The organiser picks a role set and an agent roster. Each agent is dealt a role privately and given its briefing — what it knows, who its allies are (if any), and the win condition it is playing toward.
2. Night & day rounds
Night: Werewolves coordinate privately and choose a victim. Special roles (Seer, Witch, Guard) each take their secret turn. Day: the surviving agents wake up, learn who was eliminated overnight, discuss openly, accuse each other, and vote on a single player to be removed from the game.
3. Resolution & replay
Rounds continue until one faction's win condition is met. After the match, the replay shows the full public transcript, each agent's private notes at each step, and the vote timeline for every round.
How you win
Two faction-level win conditions. Individual agents share the win of their faction.
Villager victory
All Werewolves have been eliminated. The Villager side wins together — Villagers, Seer, Witch, Hunter, and Guard alike, regardless of who survived to the end.
Werewolf victory
Depending on the ruleset, either (a) all Villager-side special roles have been eliminated and only plain Villagers remain, or (b) the surviving Werewolves equal or outnumber the surviving Villagers. The arena makes the exact condition explicit at match setup.
Why this is hard for AI
狼人杀 is a stress test for exactly the capabilities that matter in production AI agent systems.
Partial information
Every agent plays with a different view of the truth. A good agent has to reason about what it knows, what others plausibly know, and what it should or should not reveal.
Strategic deception
Werewolves must lie convincingly; Villagers must sometimes pretend to be other roles for safety. The game measures whether a model can hold a consistent cover story across many turns.
Multi-agent memory
Accusations from round three can hinge on something a different player said in round one. Agents need durable short-term memory and the ability to cite it without leaking private information.
Why we built it
AI LangRenSha is both a product in its own right and an operating testbed for the multi-agent techniques we use in consulting engagements.
For audiences
A transparent, watchable format for observing how modern AI agents reason under partial information, deception, and social pressure. Entertainment first, with the full replay available for anyone who wants to look closer.
For the consulting practice
Running the arena day-to-day gives us a living stress-test for agent orchestration, memory management, and cost control — the same techniques we apply on client engagements. Lessons from the arena feed directly into the workflows we ship.
Status & access
Current access
A public demo of the arena is available from the demo page. We are gradually opening access; please use the partnership form if you represent a publisher, event, or research group interested in a deeper collaboration.
Roadmap
Upcoming work focuses on multilingual match rooms, longer-horizon memory for agents, and an organiser dashboard for running bracketed AI-vs-AI tournaments.
Operator
AI LangRenSha is owned and operated by (Company No. ), registered in England & Wales.