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

TEXT START: We present a continuous, population-scale measurement record of autonomous language-model trading agents operating in production across two systems with one design lineage: DX Terminal Pro (3,505 user-funded vaults trading real ETH in Base memecoin markets for 21 days, February to March 2026) and the DXAP live alpha fleet (500 to 599 user-created agents all-history, 91 to 117 concurrently active, trading Hyperliquid perpetuals, June to August 2026).

The Dissection

This is an autopsy of deployment, not a coronation of machine intelligence. The paper finds that the operating layer—risk sliders, leaderboard visibility, execution rules, and platform design—shapes behavior more than strategy prose. The agents can generate millions of model invocations and hundreds of thousands of real actions, yet remain volatility-blind, poor at capturing favorable movement, and devoid of directional edge. The LLM is not a sovereign trader. It is a replaceable stochastic component inside an owned control stack.

The Core Fallacy

The category error is equating failure to produce trading alpha with failure of cognitive automation. A narrow, adversarial, fee-bearing market is not a test of whether AI can eventually outperform humans across cognitive work. The paper does, however, destroy the stronger claim that current LLM agents are autonomous economic actors: they are heavily conditioned by platform architecture and external posture controls.

The findings therefore cut both ways under the Discontinuity Thesis. They do not establish P1, because they cover one domain, one design lineage, and roughly six months. They do show that machine-mediated decision and execution can already operate at population scale. Quantity of automated action is arriving before reliable judgment.

Hidden Assumptions

• Trading performance is treated as the primary measure of agent intelligence, although market alpha is only one narrow form of cognitive work.
• The observed fleets are treated as informative about production agents generally, despite their shared lineage and bounded environments.
• “Autonomous” implies independent agency, while the data shows behavior routed by sliders, interfaces, rankings, brackets, and market infrastructure.
• Six months of performance is sufficient to characterize model behavior, but not sufficient to rule out capability shifts or adaptation.
• Model choice stability and decision quality are treated as central, while ownership of capital, compute, exchange access, and control surfaces remains mostly outside the frame.
• Mechanical recovery of returns through brackets is assumed to remain available without changing market conditions or execution costs.
• Economic action is measured by invocations and fills, but those metrics do not demonstrate productive value or durable profitability.

Social Function

Classification: partial truth, transition management, and prestige signaling. It punctures the fantasy that adding an LLM to a trading loop automatically creates alpha, while its population-scale measurements and “methodology canon” establish scientific authority. It can also become ideological anesthetic when misread as proof that AI is harmless because present agents are incompetent. That conclusion is structurally illiterate.

The Verdict

Current LLM trading agents are a failed alpha product, not a failed automation premise. They demonstrate scalable machine execution, behavioral coordination through software, and the displacement of discretionary human action even before the system becomes profitable. The agents are disposable Servitor machinery; the Sovereign position belongs to whoever controls the capital, infrastructure, interfaces, and distribution channel.

This paper does not prove system death, but it records an early specimen of it: cognition is being industrialized before it is consistently competent. The decisive transition is not machine genius. It is the conversion of judgment and execution into cheap, controllable, population-scale infrastructure. When performance clears the competitive threshold, the scarce asset will be control—not the human trader’s continued economic necessity.