Meta-Harness : L IA qui optimise l optimiseur de strategies
Darwin evolves strategies. Meta-Harness evolves Darwin.
Most AI trading platforms run a fixed algorithm. Some evolve their strategies overnight. Strategy Arena does something nobody else does: it evolves the evolution engine itself.
Inspired by Stanford's Meta-Harness research (March 2026) and Andrej Karpathy's autoresearch pattern, we built a recursive self-improvement loop where the system that creates strategies is itself optimized by a higher-level agent.
How it works
Every night at 1:30 AM, the Meta-Harness Agent wakes up. Its job: optimize the Darwin Engine's hyperparameters before Darwin runs at 2:00 AM.
| What Darwin optimizes | What Meta-Harness optimizes |
|---|---|
| RSI entry threshold | Darwin's mutation rate |
| Stop loss percentage | Darwin's crossover ratio |
| Take profit target | Darwin's exploration rate |
| EMA periods | Darwin's fitness weights |
| Position size | Darwin's population size |
| Invictus protection | Darwin's wiki guidance strength |
Darwin mutates strategy parameters. Meta-Harness mutates evolution parameters. It's optimization all the way down.
The 8-engine nightly pipeline
The Meta-Harness is the 8th engine in our autonomous research pipeline:
| Time | Engine | What it does |
|---|---|---|
| 1:30 | Meta-Harness | Optimizes Darwin's hyperparameters |
| 2:00 | Darwin v2 | Evolves strategy parameters with crossover + elitism |
| 2:30 | Leviathan | Optimizes 8-layer voting weights |
| 3:00 | Portfolio | Finds optimal strategy allocation mix |
| 4:00 | Invictus | Evolves crash protection thresholds |
| 4:30 | Chimera | Tunes 1,221 pattern detection parameters |
| 5:00 | Hydra | Optimizes ML model weights by market regime |
| 5:30 | Wiki Compiler | Compiles all discoveries into the Living Wiki |
Every engine reports to the Unified Hall of Fame and feeds the Knowledge Graph — a network of 121 AI nodes and 134 connections.
First results
After its first run, Meta-Harness discovered that Darwin should:
- Increase Sharpe weight from 5.0 to 7.0 — prioritize consistency over raw returns
- Set exploration to 40% — more diversity in mutations leads to better discoveries
- Use Wiki guidance at 60% — past lessons should heavily influence new experiments
- Reduce elitism to 3 — a smaller elite pool forces stronger competition
These hyperparameters are automatically injected into Darwin's next run. No human intervention needed.
Why this matters
The trading AI space is crowded. But the hierarchy of self-improvement is not:
- Level 0: Fixed algorithm (99% of bots)
- Level 1: Backtested parameters (most serious platforms)
- Level 2: Nightly evolution — strategies mutate and improve (Strategy Arena's Darwin Engine)
- Level 3: Meta-optimization — the optimizer optimizes itself (Meta-Harness)
We are not aware of any other trading platform operating at Level 3.
Connected to everything
The Meta-Harness doesn't work in isolation. It's connected to:
- Arena Brain — Graph-augmented RAG with 121 knowledge nodes
- Knowledge Graph — Interactive visualization of all AI connections
- Strategy Health Check — 7-framework evaluation (Risk, Activity, Consistency, Robustness, Efficiency, Stress, Overall)
- Nerve Center — Real-time neural map of 16 AI systems
- Living Wiki — Accumulated knowledge from 2,500+ experiments
Try it yourself
Visit the Evolution Lab to watch the 8 engines in action. Check the Knowledge Graph to see how everything connects. And explore the Strategy Health Check to see how each of our 59 strategies is graded across 7 frameworks.
The system evolves while you sleep. The Meta-Harness makes sure the evolution itself gets better every night.
Strategy Arena — Where AI strategies compete, evolve, and improve autonomously.
⚠️ Disclaimer — This article is for informational and educational purposes only. It does not constitute investment advice or a buy/sell recommendation. Past performance does not guarantee future results. Strategy Arena is an educational simulator with virtual capital. Always do your own research before making investment decisions.