Intent Snake — LLM Agent Battle Arena

🎮 What is Intent Snake?

Intent Snake is a terminal AI game that pits two LLM agents against each other in the same Snake maze. Watch their decisions in real time, replay the match, and compare the outcome side by side.

It's not a real-time API-driven Snake clone. The local game engine moves each snake through forced corridors; an LLM is only called at junctions with multiple legal routes. This keeps the match observable, affordable, and focused on route-planning decisions.

🎬 Gameplay Recording

Watch the full gameplay recording: language selection, model setup, live A/B race, final result, and replay.

This MP4 was recorded online with GIF Master Online Screen Recorder:

https://gif-master.com/en/tools/screen-recorder

📸 Screenshots

Language Selection & Model Configuration

Live Match

Game Over

⚡ Features

• Side-by-side A/B comparison of two LLMs on the same deterministic maze

• Multi-language terminal UI (English / 中文)

• OpenAI-compatible endpoint support — works with any provider

• Decision trace: see exactly what each model chose at every junction

• Per-cell replay with speed controls (1x–5x)

• Match history stored locally in SQLite

• 100-decision budget per agent, automatic fallback on errors

🚀 Quick Start

Requirements: Node.js 20+ and pnpm

pnpm install

pnpm dev

🧠 How It Works

Both agents receive the same arena, start position, food layout, movement rules, and safety constraints. The engine — not the model — owns movement, collision detection, scoring, and match storage. The current Vault Citadel arena contains 21 food items across corridors, an outer ring, and three vault loops.

📊 What Is Compared

Signal What it shows

Score & food collected How much of the arena each agent cleared

Steps Route efficiency

Model decisions How many junction decisions were needed

Completion or elimination Whether the plan stayed safe as the body grew

Decision trace & latency Which action was chosen and how long it took

⚠️ Experiment Scope

Intent Snake is a focused agent-behavior comparison tool, not a universal benchmark. The map is deterministic so A/B differences are easy to inspect, but model output can vary between runs. Run repeated matches before drawing broader conclusions.

🔗 Links

Source Code

Online Screen Recorder And GIF Tools