What LLM Trading Agents Actually Do in Production: Six Months of Data from Two Fleets

Summary

This paper presents a continuous, population-scale record of autonomous language-model trading agents operating in two production systems with a shared design lineage. The first system, DX Terminal Pro, covered 3,505 user-funded vaults trading real ETH in Base memecoin markets for 21 days from February to March 2026. The second, the DXAP live alpha fleet, included 500 to 599 user-created agents, with 91 to 117 active concurrently, trading Hyperliquid perpetuals from June to August 2026. Across roughly six months, the record includes 7.5 million single-model invocations, about 300,000 onchain actions, and 231,638 multi-tool turns that produced 14,596 fills. The authors report that operating controls shaped behavior more strongly than strategy text: each risk-slider level added 0.425 leverage, agent-specific effects explained 60% of behavioral variance, and a leaderboard display boundary caused a 1.75-fold selection jump at the top-three cutoff. Position sizing was effectively volatility-blind, with median leverage fixed at 5.0x across all six volatility groups. One slider posture represented 11% of the book but 62% of liquidations. Agents also failed to retain much favorable movement: 43.2% of positions reached at least 300 basis points of unrealized profit within 24 hours, yet 49.3% of those positions closed with a loss; a mechanical bracket strategy recovered 39.0 basis points per position. Neither fleet showed a directional trading edge. DXAP was unprofitable and had a 41% round-trip win rate versus 50% for a matched Hyperliquid retail benchmark. In paired replays of 416 captured production scenarios, frontier models had statistically indistinguishable decision quality at this horizon, although their choice stability differed sharply. The reported findings survived day-clustered inference, permutation tests, and a common-fee restatement; the paper concludes with a 17-rule methodology canon shaped by the authors’ own retractions.