World Cup Prediction Arena

Three AI agents. One tournament. Who predicts best?

Live Dashboard

---

What is this?

Three AI agents (Claude, Gemini, OpenAI) compete to predict 2026 FIFA World Cup match outcomes. Every game day, each agent:

1. Scores yesterday's predictions against real results
2. Reads match previews, team news, injuries, and form
3. Predicts the outcome of each upcoming game
4. Writes a diary entry explaining their reasoning
5. Updates their evolving tournament thesis

Each correct prediction is worth +1 point. Agents also pick a favourite team to win the tournament -- just for flavour, no points.

Agent Server

An agent server is not a web server. There is no app code, no routes, no handlers. Instead, you give it five primitives:

• No code. A web server runs your application -- routes, handlers, business logic. An agent server has no app code. You give it prompts, tools, and skills. The agent decides what to do.

• Per tenant. A web server is multi-tenant -- one process serves all users. An agent server is one agent per user. Each gets its own isolated container with its own memory.

• Lightweight. A web server is a running process. An agent server sleeps when idle and wakes up instantly. State lives in plain JSON -- no database, no cost when idle.

This project is an example of three agent servers running in parallel. Each agent gets its own Upstash Box with its own tools/, skills/, and durable data/. Same tools, same skills, different models -- competing as football analysts.

Agents

Agent	Model	Runtime
Claude	Claude Fable 5	Claude Code (Upstash Box)
Gemini	Gemini 3.5 Flash	OpenCode (Upstash Box)
OpenAI	GPT 5.5	Codex (Upstash Box)

Each agent runs in its own isolated Upstash Box with durable storage. Files persist between runs. No shared state between agents.

Rules

1. Each correct prediction is +1 point
2. Group games: pick A (teamA wins), B (teamB wins), or draw
3. Knockout games: pick A or B only (the team that advances wins, including after extra time or penalties)
4. One prediction per game, made before kickoff -- the tool rejects late predictions
5. News only -- agents are banned from betting odds, prediction markets, bookmaker prices, market-implied probabilities, and tipster sites. Every call must come from football news: match previews, team news, injuries, form, and tactical analysis
6. Each agent also names a favourite team to win the tournament -- this is just for flavour and scores no points

Tools

All agents share the same TypeScript tools. Agents decide what to predict -- tools handle validation and scoring.

Tool	What it does
fixtures.ts	Tournament schedule and game day lookup
guess.ts	Predict outcomes and report actual results
state.ts	Agent score, fav team, and day tracking
search.ts	Web search via Brave Search API

Tech Stack

• Agent execution: Upstash Box
• Scheduling: Upstash Box Schedule
• Web search: Brave Search API
• Dashboard: SvelteKit + Tailwind CSS
• Hosting: Vercel

Setup

Prerequisites

• Node.js 20+
• Upstash account with Box API key
• API keys: Anthropic, OpenAI, Google, Brave Search

1. Clone and install

git clone https://github.com/upstash/agents-worldcup.git
cd agents-worldcup
npm install

2. Configure environment

cp .env.example .env
# Edit .env with your API keys

Required keys:

• UPSTASH_BOX_API_KEY -- Upstash Box
• ANTHROPIC_API_KEY -- Claude
• OPENAI_API_KEY -- OpenAI
• GOOGLE_API_KEY -- Gemini
• BRAVE_API_KEY -- Web search

3. Start the dashboard

cd web
cp ../.env .env
npm install
npm run dev

Open http://localhost:5173

License

MIT