OpenReward

complex-worlds-hackathon-games

Khalid

Khalid/complex-worlds-hackathon-games

Decision-Making in GamesMultimodal ReasoningAgentic AI Evaluation
API Endpoint
Leaderboard
Loading leaderboard...
README

rl-env-hackathon-complex-worlds

OpenReward agentic environments for complex games.

Overview

This project provides OpenReward-based environments that enable LLM agents to play games including Settlers of Catan, Overcooked, and Diplomacy. Each environment exposes game state and actions through a tool-based interface, allowing agents to reason about and interact with the game world.

The project includes a trajectory analysis system that learns from past games to build experience libraries, which can be injected as system prompt extensions to improve agent performance.

Synthesis Pipeline

Synthesis Pipeline

Features

  • Three Game Environments: Catan, Overcooked, and Diplomacy, each with configurable difficulty levels
  • OpenReward Integration: Tool-based environment interface designed for LLM agents
  • Claude Agent SDK Support: Uses your Claude subscription for authentication (no ANTHROPIC_API_KEY required)
  • Experience Library System: Analyze past trajectories and extract learned strategies
  • Visualization Support: Catan includes live web visualization of game state
  • Testing Suite: Scripted smoke tests and end-to-end Claude agent tests for each environment

Installation

Requires Python 3.10+.

git clone https://github.com/yourusername/rl-env-hackathon-complex-worlds.git
cd rl-env-hackathon-complex-worlds
pip install -e .

Project Structure

src/ catan_env/ # Settlers of Catan environment overcooked_env/ # Overcooked AI environment diplomacy_env/ # Diplomacy environment scripts/ # Sweep, analysis, and visualization scripts tests/ # End-to-end and scripted tests trajectories/ # Saved game trajectories (JSON) experiences/ # Learned experience libraries (Markdown)

Environments

Catan (src/catan_env/)

Settlers of Catan environment powered by Catanatron.

Difficulty Levels:

  • very_easy: 1 RandomPlayer, 5 VPs to win, friendly robber
  • easy: 3 RandomPlayers, 8 VPs to win, friendly robber
  • medium: 3 WeightedRandomPlayers, 10 VPs to win, hostile robber
  • hard: 3 ValueFunctionPlayers, 10 VPs to win, hostile robber
  • very_hard: 3 AlphaBetaPlayers, 10 VPs to win, hostile robber

Available Tools:

  • get_state() - Return a textual summary of the current board state
  • list_legal_actions() - Return numbered list of legal actions for the current decision
  • play_action(index: int) - Apply a legal action by index and advance the game

Overcooked (src/overcooked_env/)

Overcooked AI environment for cooperative cooking challenges.

Diplomacy (src/diplomacy_env/)

Diplomacy environment for strategic negotiation and conquest.

Khalid/complex-worlds-hackathon-games | OpenReward