IMC Prosperity 4: A Multi-Agent Trading System

10 AI agents walk into a trading competition. They argue, backtest, break each other's code, and try to win $50K.

This is an experiment in multi-agent orchestration for IMC Prosperity 4, a global algorithmic trading competition for university students. Instead of one person writing one algorithm, we built a system where 10 specialized Claude Code agents with distinct personalities collaborate (and clash) to develop trading strategies.

The Agents

Agent	Personality	Job
Quant	Academic, citation-heavy, distrusts heuristics	Research papers, mathematical models, theoretical backing
Market Maker	Talks like a trader, profit-obsessed	Spread optimization, fair value, position management
Stat Arb	Skeptical data scientist ("is this edge real?")	Basket arb, pairs trading, z-score mean reversion
Options	Precise, methodical, obsessed with Greeks	Black-Scholes, vol modeling, delta hedging
Backtester	Completely neutral, numbers only	Runs backtests, maintains strategy leaderboard
Red Team	Paranoid contrarian with veto power	Bug hunting, overfitting detection, adversarial review
Manual Solver	Puzzle nerd, shows all work	Game theory, optimization for manual challenges
Copycat	Competitive intel, studies the winners	Scans GitHub for top teams' strategies
Playwright	Efficient browser operator	Platform scraping, leaderboard, algo submission
Repo Maintainer	Documentarian	Keeps this README alive and the repo sharp

How It Works

You: "/prosperity new-round 2"
                |
    +-----------+-----------+
    |           |           |
  Quant    Market Maker   Copycat    (research in parallel)
  Stat Arb   Options    Manual Solver
    |           |           |
    +-----------+-----------+
                |
           Backtester         (test everything)
                |
            Red Team          (break everything)
                |
         Final trader.py      (merge the winners)

Each agent writes proposals to its own workspace in agents/. The backtester maintains a leaderboard. The red team has veto power. Agents share a "brain" (.claude/shared-knowledge/BRAIN.md) for collective knowledge and can message each other directly in Agent Team mode.

Competition

IMC Prosperity 4 (April 14-30, 2026)

• 5 rounds, 16 days
• Algorithmic trading (Python bot) + manual challenges (game theory puzzles)
• $50,000 prize pool
• ~13,000+ teams globally

Project Structure

imc-prosperity-2026/
  .claude/
    agents/              # 10 agent definitions with distinct personalities
    shared-knowledge/    # Shared brain (BRAIN.md)
    skills/prosperity/   # /prosperity skill (orchestrator)
  agents/                # Agent workspaces (proposals, results, reviews)
    quant/proposals/
    market-maker/proposals/
    stat-arb/proposals/
    options/proposals/
    backtester/leaderboard.md    # Strategy rankings
    red-team/kill-list.md        # Vetoed strategies
    copycat/registry.md          # Public submission database
    manual/solutions/
  src/
    trader.py            # The submitted algorithm
    strategies/          # Per-strategy modules (merged at submit time)
  docs/                  # Competition docs, wiki scrapes, resources
  data/                  # Round data capsules
  backtests/             # Backtest results

Quick Start

# Install backtester
pip install -U prosperity3bt

# Run backtest (once you have a strategy)
prosperity3bt src/trader.py 1

# With visualizer
prosperity3bt src/trader.py 1 --vis

Round Log

Updates after each round with results, agent disagreements, and lessons learned.

Tutorial Round

Status: Setting up. Agents deployed. Products: EMERALDS (stable, limit 80), TOMATOES (volatile, limit 80).

---

Built with Claude Code by @aflekkas