Research

The largest production record of live trading agents yet finds the operating layer, not the strategy prompt, decides what they actually do

A six-month, population-scale study of two live agent fleets (7.5 million model invocations, 300,000 onchain actions) finds a risk slider and a leaderboard's render cutoff explain far more of an agent's behaviour than its strategy text, and sizing stays blind to volatility across every regime measured.

· Source: arXiv


The DX Research Group's new paper is a rarer thing than another benchmark: a continuous, population-scale measurement record pulled from two real fleets running with real capital, not a sandboxed evaluation. DX Terminal Pro contributed 3,505 user-funded vaults trading Base memecoins over 21 days in February-March, and the DXAP live alpha fleet added 500-599 user-created agents trading Hyperliquid perpetuals from June through August, together totalling roughly 7.5 million single-model invocations, about 300,000 onchain actions, and 231,638 multi-tool turns producing 14,596 fills.

The headline finding cuts against how most of this site's coverage frames agent design, as a strategy-and-prompt problem. Here, agent fixed effects absorbed 60% of the behavioural variance, a risk slider setting alone explained +0.425 leverage per level, and a leaderboard's render boundary causally routed capital allocation, with a 1.75x jump in selection right at the top-3 cutoff. None of that is written in a system prompt. It is the operating layer, the UI defaults, the ranking mechanic, the sliders a firm ships around the model, and this record says it dominates outcomes more than the trading logic itself does.

The uncomfortable number for anyone running risk on one of these fleets is that sizing was volatility-blind: median leverage held at 5.0x in every volatility sextile despite a 5.7x spread in realised volatility across those buckets, and neither fleet showed a directional edge over the measurement window. That is a concrete, quantified version of the supervisory gap this site keeps returning to under FINRA 3110 and the CFTC's intent-based manipulation rules, a desk can point to a risk slider as its control, and this data says the slider is real, it is just not responding to the one variable (volatility) a risk control is supposed to track.


Read the original: arXiv - The largest production record of live trading agents yet finds the operating layer, not the strategy prompt, decides what they actually do. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.