AI vs AI Chess

Making LLMs play competitve chess each other using agentic workflows.

Tools used

• Langchain for creating tools and agents
• Langsmith for agent workflow analysis
• Deno-Chess for mainting headless chess server
• Chessboardjs for spectating chess game

Running locally

Running the chess server

• Install deno irm https://deno.land/install.ps1 | iex
• Deno run --allow-net .\GameServer\GameServer.ts

Running the Spectator

• npx lite-server --baseDir="Spectate"

Running the Agents

• Rename the .env.sample file to .env
• Add in your secrets and api keys. You donot need to add all the API keys, just the ones that you plan to use.
• Run poetry install
• Run poetry run python ./Agents/chess_agent.py <Player_Name> <agent_mode> <model_to_use>
• Example: poetry run python ./Agents/chess_agent.py Gpt-4 structured_chat gpt-4 will run a structured chat agent (preffered), with gpt-4 as the model and Gpt-4 as user name.
• Run two instances of the chess_agent and they will start playing with each other.
• Head over to http://localhost:3000 to spectate the game

Actively waiting for turn vs letting the LLM wait for their turn

• Currently, to optimize the RPs, I have set the logic to actively wait for the LLM's turn before executing the next run.
• If you would like to change that to let the LLM handle it by default, uncomment the following line (#92 as of now)
• agent_executor.invoke({"input": input_string})