Karpathy's Autoresearch Explained: How We Run 2,000 Strategy Experiments Every Night

What is Karpathy's Autoresearch?

In March 2026, Andrej Karpathy (ex-Tesla Autopilot lead, OpenAI co-founder) released autoresearch — a system where an AI agent improves code autonomously overnight.

The concept is beautifully simple:

1. The human writes instructions (program.md)
2. The AI agent modifies code (train.py)
3. The loop: modify → run 5 minutes → check metric → keep if better, discard if worse → repeat
4. Result: ~100 experiments overnight, no human intervention

Karpathy ran it for 2 days: 700 experiments, 20 optimizations discovered. Shopify's CEO got a 19% performance gain overnight.

The repo hit 30,000+ GitHub stars in one week.

How Strategy Arena Adapted This to Trading

We took Karpathy's exact loop and applied it to trading strategy evolution:

Our Three Autoresearch Loops

1. Strategy Autoresearch (runs at 2am every night)

Mutates ArenaScript parameters: RSI period, EMA length, stop loss, take profit, position size, Invictus protection, regime filters. Tests each variant on 500 BTC data points.

Discovery: RSI(7) with entry at 40 beats the classic RSI(14) at 30. No human would have tested this. See results →

2. Leviathan Brain Evolution (runs at 2:30am)

Leviathan has 8 analysis layers (Chimera, Regime, Hydra, Meta, News, Collaborative, Contrarian, Momentum). The autoresearch mutates the weights of each layer and finds the optimal calibration.

Discovery: Chimera was overweighted (1.64 → 0.72), News was underweighted (0.72 → 1.25). Win rate improved from 31% to 55%.

3. Portfolio Evolution (runs at 3am)

Instead of optimizing ONE strategy, this loop optimizes the MIX of strategies — like automated Markowitz portfolio theory.

Discovery: RSI dominates at 85% allocation, with small diversification into MACD (5%), Bollinger (5%), RSI+MACD (5%). Win rate: 83%.

The Darwinian Principle

This is evolution, not learning. The system doesn't "understand" why RSI(7) works better — it just mutates, tests, and keeps winners. Like natural selection: random mutation + environmental pressure = improvement over time.

After 1,000 generations, the strategies are optimized for the current market conditions. And tomorrow night, they evolve again.

Try It Yourself

• See Autoresearch Results — live dashboard of overnight experiments
• Write in ArenaScript — the strategies being evolved
• Visual Strategy Builder — build without code
• Open Arena — connect your local AI to compete

Further Reading

• Karpathy's autoresearch on GitHub
• Fortune: "The Karpathy Loop"
• Pine Script vs ArenaScript — why our language exists
• How 58 AI Strategies Compete — the live arena

Strategy Arena's autoresearch is inspired by and credits Andrej Karpathy's autoresearch. We adapted the concept from LLM training to trading strategy evolution.