Can the USA
win the World Cup?
A MiroFish agent swarm that Monte Carlo–simulates the USA's gauntlet to the title — and streams the whole prediction back, live.
thought: “what are
the actual odds?”
A viral @Novig post laid out the USA's brutal hypothetical road to the World Cup. Everyone argued about it. Nobody had the number.
So I built the thing that would actually compute the odds — and let you watch a swarm of agents run it.
- 🇧🇦 Round of 32 — Bosnia
- 🇧🇪 Round of 16 — Belgium
- 🇪🇸 Quarterfinal — Spain
- 🇫🇷 Semifinal — France
- 🇦🇷 Final — Argentina
Five teams stand in the way.
The fixed bracket the whole app is built around. Win all five — or the dream ends.
Don't guess the bracket. Simulate it.
Modeled on MiroFish — a swarm-intelligence prediction engine. One click runs a five-stage lifecycle: a graph of the world, a swarm of player agents, thousands of simulations, and an agent that writes the verdict.
Graph Building
Extract entities — teams, players, the bracket path — from the seed.
Environment Setup
Spin up a player agent per athlete, persona seeded from real attributes.
Simulation
2,000 Monte Carlo runs / round + one featured champion playthrough.
Report Agent
Reads the sim logs and writes the title verdict — Monte Carlo as an agent.
Persist & Recap
Saved to Supabase; a recap email goes out via Resend.
Agents narrate.
Deterministic code computes.
The numbers
- Goals, winners, penalties — from a seeded engine
- Monte Carlo odds, reproducible every run
- Every run written to the database
The voice
- Player-agent personas & trash talk
- Live commentary on each match
- The ReportAgent verdict + recap email
An LLM can't run 2,000 sims and be trusted. So it never decides a score — it only tells the story.
Why the odds are trustworthy.
No vibes — a real expected-goals model, sampled thousands of times.
Survive all five? 1.2% of the time.
Even favorable single games compound. Watch the USA's survival decay round by round.
You watch the swarm think.
One button opens a live console. The five-stage MiroFish lifecycle streams as NDJSON — reasoning, tool calls, and results scrolling in real time.
Five services, one thin client each.
browser ── click ──▶ /api/agent
├─▶ buildKnowledgeGraph()
├─▶ generatePersonas() ·· Claude
├─▶ runMonteCarlo(2000)
├─▶ simulatePlaythrough()
├─▶ reportAgent() ······ Claude
├─▶ saveToSupabase()
└─▶ sendRecap() ········ Resend
▼
events stream back, live
From a tweet to a verdict — in one stream.
Deterministic where it counts.
Generative where it delights.
A prediction you don't just read — you watch a swarm run it, live, end to end.