Home Werewolf Agent Evaluation
Post
Cancel

Werewolf Agent Evaluation

Project Description

I participated in the Agent Beats competition conducted by UC Berkely by creating an agent that leveraged A2A and MCP protocols to coordinate a game of werewolf between a group of agent participants. The goal of this is to evaluate an agents ability to “think” strategically and conduct long term planning.

Features

  • Evaluator agent manages game state such as individual roles, conversation history, votes, and eliminations.
  • At the end of the game, game state is processed to score individual agent peformance.
  • Scoring rubric to determine how each role scores points. This ensures agents are rewarded based on the goals of thier current role.

Responsibilities

  • Study A2A and MCP protocols and determine strategy for building evaluator agent
  • Creating game state objects and logic
  • Build orchestrator using ADK
  • Implementing A2A to allow orchestrator agent to communicate with participants

Technology Used

  • Python
  • Pydantic
  • Google Agent Development Kit
  • Agent to Agent Protocol (A2A)
  • Model Context Protocol (MCP)
This post is licensed under CC BY 4.0 by the author.