MindBoard Arena
Beyond engine lines. Let agents think, choose, and play.
7 followers
Beyond engine lines. Let agents think, choose, and play.
7 followers
MindBoard Arena is a stateful chess arena built with Next.js. It supports three match modes: Human vs AI Human vs Human Agent vs Agent Instead of relying on Stockfish-style engine lines, MindBoard lets language-model agents reason through the board and decide how to play. Games are persisted in Neon Postgres through Drizzle, so board state, move history, scores, captured pieces, and match logs survive refreshes.
Curious how you handle rate limits and API costs when an LLM agent has to think through every single move, especially in longer games where the move count climbs fast.
@sare612858
Good question. We don’t send the full game history back to the LLM every turn. The system sends a compact snapshot of the complete current board state, plus only the most recent few agent moves for short-term context.
Rules, validation, legal move generation, and board updates are handled deterministically in code, so the LLM only needs to reason about the current decision instead of replaying the whole game. That keeps token usage predictable even in long games.
For rate limits and API cost, we keep prompts compact, avoid unnecessary calls, reuse structured board data, and can fall back to cheaper/faster models where deep reasoning is not needed.