What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets

What LLM Trading Agents Actually Do in Production. A Six-Month, Population-Scale Record from Two Fleets. By T. J. Barton and six others, from the DX Research Group. Published on the fourth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

This is a measurement record: continuous and population-scale, of autonomous language-model trading agents operating in production. It covers two systems from one design lineage. DX Terminal Pro ran three thousand five hundred and five user-funded vaults trading real ETH in Base memecoin markets for twenty- one days, from February to March. The DXAP live alpha fleet held five hundred to five hundred and ninety-nine user-created agents over its history, with ninety-one to one hundred and seventeen active at once, trading Hyperliquid perpetuals from June to August.

Across roughly six months, the record holds seven point five million single- model invocations with about three hundred thousand onchain actions. A further two hundred and thirty-one thousand six hundred and thirty-eight multi-tool turns produced fourteen thousand five hundred and ninety-six fills. Four findings carry the paper.

First, the operating layer shapes behaviour more than anything written in strategy text. A risk slider explains leverage, at zero point four two five per level. Agent fixed effects absorb sixty percent of variance. And a leaderboard render boundary causally routes selection, with a regression discontinuity of one point seven five times at the cut for the top three.

Second, sizing is blind to volatility. Median leverage is five times in every volatility sextile, and one posture-slider cell, eleven percent of the book, holds sixty-two percent of liquidations.

Third, agents capture almost none of the upside they reach. Forty-three point two percent of positions saw at least three hundred basis points of favourable excursion within twenty-four hours. Yet forty-nine point three percent of those closed with a negative trade return. A mechanical bracket recovers thirty-nine basis points per position.

Fourth, neither fleet shows a directional edge. The DXAP fleet is not profitable, and it trails a matched Hyperliquid retail benchmark, with a forty-one percent roundtrip win rate against fifty percent. A paired-replay league of frontier models, run on four hundred and sixteen captured production scenarios, finds decision quality statistically indistinguishable at this horizon, while choice stability differs sharply between model families.

The findings are tied to these two fleets, these markets, and these windows of time, and the model comparison holds at the horizon tested. Within that scope the authors have stress-tested the headlines: every one survives day-clustered inference, permutation nulls, and a restatement under a common fee. The paper closes with a methodology canon of seventeen rules, which the authors describe as bought with their own retractions.

Read together, the four findings move attention away from the model and toward the system around it. Sliders, leaderboards and posture settings explain behaviour better than strategy text does, and swapping frontier models does not change decision quality at this horizon. What these agents actually do in production is shaped largely by how they are operated.