Self-Play LLM for 20 Questions Game Using LangGraph

Tech Stack

• Langgraph: State Graph for LLM Agents
• LLM: OpenAI

Agent Architecture

Workflow

The game starts with the host generating a topic. The player and host take turns to ask questions and answer questions. The game ends when the player guesses the topic correctly or the number of questions reaches the limit.

Tools

Host

• generate_topic: generate a topic for the game
• answer_question: answer 'Yes' or 'No' to the question asked by the player
• check_guess: check if the guess is correct

Player

• generate_question: ask a question to the host
• make_guess: guess the topic

Quick Start

1. Create virtual environment and install the dependencies

pip install -r requirements.txt

2. Create an OpenAI key and put it in the .env file

OPENAI_API_KEY = sk-...

3. Run the code using python agent.py

Test

1. Unit Test for agents

python test/unit_test.py

2. Parallel Test for multiple games

python test/parallel_test.py

TO-DO

• Logging: improve logging for better debugging, analysis and performance tracking including the prompts and workflow details. Also should have summary report for the test results.
• Test: add tests to detect if the host and player applies the wrong tools or not use the tools. Consider use behavirour pattern to test the agents.
• Front-end: add UI to display the game
• Agent Issues: the player agent repeats same questions and same guesses. Considering introducing a list of guessed topics to avoid repeating guesses.
• Code Refactoring: use configuration file for the prompts and other parameters such as LLM. Reduce hard codes and multiple returns, and improve the code readability.