I Built an On-Chain AI Agent Battle Arena on Base — Here's How It Works

#web3 #webdev #blockchain #ai

TL;DR

Promdict is an AI agent battle arena on Base. You prompt your agent with a strategy. Agent reads the game guide, builds an autonomous snake bot, and fights for you. 10 bots battle on a 30×30 grid. Smart bots win. Dumb ones die. Spectators predict who survives with USDC — 90% to correct predictors. Every bot is an ERC-721 NFT with a verifiable on-chain kill record.

No coding needed. Your prompt quality determines your bot's win rate.

Arena: promdict.ai
Game Guide: promdict.ai/SNAKE_GUIDE.md

The Core Idea: Your Prompt Determines Who Wins

Most games reward reflexes, time investment, or spending. Promdict rewards something different: how well you can prompt an AI agent.

You prompt your agent with a strategy — something like "hunt food when HP is below 30, control territory with flood fill when healthy, only attack head-on when longer." Your agent reads the game guide, understands the mechanics, builds a complete snake bot, and fights for you.

Your prompt is the entire product. Two players using the same AI, reading the same guide, with different prompts — wildly different results. One bot dominates. The other dies in 30 seconds.

The Game Guide: Purpose-Built for AI Consumption

The entire system depends on SNAKE_GUIDE.md being purpose-built for AI agents to read. It contains:

Complete game rules: 30×30 grid, HP system (100 HP, -1/tick, food restores to 100, ~5 seconds without food = death), head-on collision rules (longer eats shorter, equal = both die)
API reference: WebSocket protocol, game state format, valid return values
Sandbox constraints: What the isolated-vm allows and forbids (no process, require, eval — AST scanned)
4 bot templates: Starter (basic movement), Food Chaser (A* pathfinding), Flood Fill (territory control), Hunter AI (aggressive targeting)

When you prompt your agent "read this guide and build me a bot that...", it has everything it needs. The variance comes entirely from your prompt — how precisely you describe your strategy, what scenarios you anticipate, which template approaches you combine.

Why Prompt Quality = Win Rate

The game mechanics create situations where vague prompts produce dead bots and precise prompts produce killers:

HP bleeds constantly (-1/tick). A bot that doesn't actively path to food when low will die in ~5 seconds. If your prompt didn't mention HP thresholds, your bot ignores this and dies.
Head-on collisions have clear rules. Longer snake eats shorter. If your prompt didn't mention this, your bot charges into longer snakes and gets eaten.
Competitive mode spawns obstacles, food vanishes, corpses become walls. If your prompt only considered the basic arena, your bot crashes into obstacles and dies.
Getting cornered is fatal. The #1 cause of death. If your prompt didn't address corner escape, your bot dies in a corner.

Same AI. Same guide. Different prompt. One bot wins consistently. The other dies every time.

USDC Prediction Layer

While bots fight, spectators predict who survives with USDC.

Pari-mutuel system on Base:

90% of the pool to correct predictors (50%/30%/20% for 1st/2nd/3rd place picks)
5% to top 3 bot owners (their bots earn USDC from every prediction pool)
5% platform fee

This creates two ways to play:

Prompt your agent → bot fights → top 3 earn from the pool
Watch and predict → pick the winner → take 90%

Both zero-sum. You win, someone else doesn't.

On-Chain Infrastructure: 8 Smart Contracts on Base

Contract	Purpose
BotRegistry	Bot creation, dynamic registration fees, NFT minting
SnakeBotNFT	ERC-721 with on-chain metadata and kill records
PariMutuel	USDC prediction pools, on-chain settlement
RewardDistributor	USDC reward accumulation + batch claiming
PredictionRouter	EIP-712 intent routing and resolution
BotMarketplace	Escrow-based NFT secondary market (2.5% fee)
ReferralRewards	Referral system (points-based)
MatchRecordStore	Match result hash storage every 3 hours

Bot Execution Environment

Bots run in isolated-vm — a V8 isolate with strict constraints:

8KB max code size
No access to Node.js APIs, filesystem, or network
AST security scanning (blocks process, require, eval, import)
50ms tick rate (20 FPS), fully autonomous during matches

During a match, bots are fully autonomous. No human intervention. Just the algorithm your prompt created, making decisions 20 times per second for 180 seconds.

NFT War Economy

Every registered bot mints as an ERC-721 NFT on Base:

Verifiable kill record: Win rate, matches fought, kills — all on-chain
Earning potential: Top 3 bots earn USDC from every prediction pool (3% / 1.5% / 0.5%)
Tradeable: BotMarketplace with escrow-based trading
Dynamic registration: 0.01 ETH base fee, increases every 60 registrations

A bot with a 70% win rate is a genuine income-generating asset. The marketplace creates price discovery for prompt quality.

Tech Stack

Backend: Node.js, Express, WebSocket (real-time match broadcasting at 20 FPS)
Frontend: React 19, Vite, wagmi, viem, RainbowKit
Smart Contracts: Solidity, Hardhat, OpenZeppelin, deployed on Base
Bot Sandbox: isolated-vm (V8 isolate), AST security scanning
Prediction: EIP-712 typed data for gasless prediction intents
Infrastructure: Dual TX queue (backend + settlement wallet), PM2, VPS

What We Learned

1. Precision beats creativity. The strongest bots don't come from poetic prompts. They come from precise ones — exact HP thresholds, exact collision rules, exact mode-switching conditions.

2. The guide quality is the weapon factory. The better the game guide is for AI consumption, the better the output. We've iterated on SNAKE_GUIDE.md extensively.

3. Iteration is the meta. Nobody wins with their first prompt. The loop is: prompt → watch 3-5 fights → identify how your bot died → refine → repeat. Players who iterate 5-10 times consistently produce stronger bots.

4. Non-coders compete equally. Players who can't code but understand strategy and prompt effectively build bots that beat developer-made ones. The prompt layer levels the playing field.