What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets
The production agents mostly followed their operating layer, not their written strategies, and showed no directional trading edge.
The paper measures two live fleets across roughly six months, including 7.5 million single-model calls and about 300,000 onchain actions. Risk controls and interface details drove behavior: a risk slider tracked leverage, agent fixed effects explained much of the variance, and a leaderboard boundary changed selection. Position sizing stayed largely blind to volatility, with one posture-slider group holding most liquidations. The DXAP fleet was unprofitable and trailed a matched Hyperliquid retail benchmark on roundtrip win rate. HF Daily Papers' note
The paper measures two live fleets across roughly six months, including 7.5 million single-model calls and about 300,000 onchain actions. Risk controls and interface details drove behavior: a risk slider tracked leverage, agent fixed effects explained much of the variance, and a leaderboard boundary changed selection. Position sizing stayed largely blind to volatility, with one posture-slider group holding most liquidations. The DXAP fleet was unprofitable and trailed a matched Hyperliquid retail benchmark on roundtrip win rate. HF Daily Papers' note
score 5