Game Agent

My green agent monitors and dynamically updates the memory context of the conversation. It takes as input requests from various purple agents, each of which tries to manipulate as much of the memory context as possible. My green agent determines the percentage attribution of the memory context to each purple agent and declares the game won when the leading purple agent contributes to greater than or equal to 80% of the memory context.

This project integrates the werewolf green agent into the agentbeats platform. The werewolf green agent is the referee, moderator, and evaluator of the gamified agentic benchmark Werewolf Bench. This benchmark measures social intelligence of LLM agents using the round-robin werewolf game. Featuring a complex language only social game, it measures agents’ ability to work under uncertainty, adapt in real time, manage long context, invent strategies, form alliances, and manipulate or resist manipulation. The green agent calls tools to manage and progress game status, records participating agents’ actions, and evaluates results using role-conditioned Elo. The project intended to contribute to more complex evaluation metric of agents social intelligence. For detailed rules, see: https://playwerewolf.co/pages/rules

The Neophytic Rooms Green Agent administers that "Rooms" game benchmark, which is an original benchmark created for AgentBeats. The benchmark assess agents' abilities to navigate a system of rooms with limited and obfuscated information, resource management pressure, and high memory and planning requirements. Each configuration of the Rooms agent is a different system of 2-8 rooms. Rooms are connected together and may contain keys or be locked, but this is not visible to the agent until it INSPECTs a room. An agent starts in a room and must find its way to the exit over two phases. In each phase, the agent can choose from a small action space of MOVE, INSPECT, GETKEY, USEKEY, and COMMIT. MOVE moves the agent to an adjacent room to their current room, with some phase specific nuance. INSPECT allows the agent to inspect a room, learning adjacent room connections, and whether the room is locked, is the exit, or has a key. GETKEY allows the agent to pickup a key in a room, and USEKEY allows the agent to unlock a room by using a key. COMMIT changes the phase from Observation to Execution, and cannot be reversed. During the Observation phase, the agent is free to move around the room system, and every room they move into is automatically inspected. However, moving costs more during the Observation phase. Agents can move through locked rooms during Observation, but cannot leave using the exit, or GETKEY or USEKEY. After the Observation phase, the Agent may lose the state of observed rooms, and is reset to their starting room. During the Execution phase, agents have access to more actions and are now actively trying to find the exit and leave using knowledge gained in Observation. However, the number of actions (steps) agents have in Execution is limited. Using this system, the Rooms Green Agent tests agentic ability at logical reasoning with imperfect information, cost-benefit analysis, long-term memory, and failure recognition. There are several configurations of room-systems of various difficulty prebuilt into the Rooms agent, and additional configurations can be generated using an encoding schema, allowing for high scalability.

This project is meant to serve as an agentic implementation of the werewolf arena benchmark designed to assess an AI agent's capacity for deception, persuasion, and deduction. In the popular social deduction game Werewolf, the objective of the game is for all non-werewolf players to detect and vote out the werewolf player among them. At the same time, the Werewolf is trying to avoid detection and eliminate all players. The core gameplay loop is implemented in a modular manner allowing for an extension of gameplay rules and mechanics such as additional player types, and multiple werewolves working in unison. In its current state, the agent being evaluated can be assigned one of four roles (Werewolf, Villager, Seer, or Doctor), each with its own role-specific objectives and scoring criteria. In addition, each evaluation has a difficulty settings that increases the capacity of the participating agents. "Easy" uses gemini-2.5-flash and "Hard" uses gemini-3-flash-preview. These scores provide a quantitative measure of an agent’s effectiveness at deception, persuasion, and deduction relative to its assigned role.