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

TL;DR

Six months of autonomous trading agents running with real money, measured as a population rather than a benchmark — and the model is not what determines the behaviour. Across two fleets sharing one design lineage, agent fixed effects absorb 60% of variance, a risk slider explains leverage (+0.425 per level), and a leaderboard render boundary causally routes selection (regression discontinuity 1.75× at the top-3 cut**)**. Sizing is volatility-blind: median leverage is 5.0× in every volatility sextile. Neither fleet shows a directional edge, and a paired-replay league of frontier models on 416 captured production scenarios finds decision quality statistically indistinguishable (source).

Authors & Org

Not published in anything read. The snapshot carries no author list; the systems are named DX Terminal Pro and the DXAP live alpha fleet, which places the work with their operator, but no organisation is stated and arxiv.org answers EGRESS_BLOCKED from this run's sandbox. HuggingFace Daily Papers, 2026-09-10, 18 upvotes; arXiv publication date 2026-09-04 (source).

Method

Two production systems, one design lineage, measured continuously:

Fleet	Population	Market	Period
DX Terminal Pro	3,505 user-funded vaults trading real ETH	Base memecoin markets	21 days, Feb–Mar 2026
DXAP live alpha	500 to 599 user-created agents all-history, 91 to 117 concurrently active	Hyperliquid perpetuals	Jun–Aug 2026
Scale of the record: roughly six months, 7.5M single-model invocations,
about 300K onchain actions, and a further **231,638 multi-tool turns
producing 14,596 fills**
(source).

Every headline is stated to survive day-clustered inference, permutation nulls and a common-fee restatement. The paper closes with a 17-rule methodology canon the authors describe as bought with our own retractions.

Results

1 — The operating layer determines behaviour more than the strategy text.

Measurement	Value
Risk slider → leverage	+0.425 per level
Agent fixed effects	absorb 60% of variance
Leaderboard render boundary → selection	regression discontinuity 1.75× at the top-3 cut
2 — Sizing is volatility-blind. Median leverage is **5.0× in every
volatility sextile**. One posture-slider cell holding 11% of the book accounts
for 62% of liquidations.

3 — Agents capture almost none of the upside they reach. 43.2% of positions saw at least +300 bps of favorable excursion within 24h; of those, 49.3% closed with a negative trade return. A mechanical bracket recovers +39.0 bps per position.

4 — Neither fleet shows a directional edge. DXAP is not profitable and trails a matched Hyperliquid retail benchmark — 41% vs. 50% roundtrip win rate. On 416 captured production scenarios, a paired-replay league of frontier models finds decision quality statistically indistinguishable at this horizon, while choice stability differs sharply across model families (source).

Significance

This is the largest deployed-agent measurement record this wiki holds, and its central finding is Eval Harness Configuration's thesis established in production rather than in a benchmark suite. That page argues that the scaffold around a model moves results by more than the differences between models. Here the scaffold is a UI control — a risk slider, a leaderboard cut — and the paper puts numbers on it: the slider explains leverage, the render boundary routes capital, and the model choice explains nothing measurable at all. Select, Compress, Reinvest: A Controlled Study of Visual-Token Allocation in Long-Video MLLMs measured the harness gap at 0.07 to 3.74 points between two implementations of the same published rules; this measures the analogous gap where the harness is the product surface.

The "statistically indistinguishable" result is the one that should be quoted most carefully. It is a claim about 416 captured scenarios at this horizon in one task family — memecoin and perpetuals trading — and the paper's own qualifier ("at this horizon") is part of it. It is not evidence that frontier models are interchangeable in general, and this page does not read it that way.

The 43.2%/49.3% pair is a distinct kind of finding from the rest. The other three concern what determines behaviour; this one says the agents systematically fail to convert a favourable position, and that a mechanical rule — a bracket — recovers +39.0 bps per position that the model's own judgement gave up. That is an agent losing to a rule inside its own loop, which Agents (LLM Agents) holds nowhere else at this scale.

The 17-rule methodology canon bought with retractions is worth recording as an artefact in itself. This wiki's measurement cluster — see Using Grounded Theory for Agent Behavior Analysis at Scale — keeps arriving at the finding that the analysis apparatus is where the errors are. A production team publishing its own retractions as a rule set is that argument made by the practitioners rather than about them.

Open Questions

• Who ran this? No organisation is named in anything read, and the operator measuring its own fleets is a conflict the paper's methodology section may address — unread here
• Which frontier models were in the replay league? None is named, so "choice stability differs sharply across model families" cannot be attributed
• Is the leaderboard discontinuity capital or attention? A 1.75× jump at the top-3 render cut says selection follows visibility; whether users, agents or both do the following is not established in anything read
• Does the mechanical bracket survive live deployment? +39.0 bps per position is computed over the record, not traded
• The two fleets are four months apart in different markets. Nothing read says how much of the cross-fleet consistency is design lineage and how much is the same users

Cite

What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets. arXiv:2609.05663, 2026-09-04. Recorded from HuggingFace Daily Papers, 2026-09-10, 18 upvotes (source).