Does AI trading actually work?

What we ran

DX Terminal Pro was a 21-day deployment in which 3,505 user-funded agents traded real ETH in a bounded onchain market on Base. Users wrote strategies and set controls; only the agents could trade. It produced 7.5 million agent invocations, roughly 300,000 onchain actions, about $20 million in volume, more than 5,000 ETH deployed, and 99.9% settlement success for policy-valid submitted transactions. The full write-up is on arXiv: Operating-Layer Controls for Onchain Language-Model Agents Under Real Capital.

What decided the outcome

The base model on its own did not make agents reliable. What did was the layer around it: how the strategy is compiled into a prompt, typed controls, policy validation before execution, execution guards, memory design, and a trace of every turn. Pre-launch testing surfaced failures that text-only benchmarks miss: fabricated trading rules, fee paralysis, numeric anchoring, cadence trading, and misread tokenomics.

• Fabricated sell rules fell from 57% to 3% of sell decisions after targeted harness changes.
• Fee-led no-trade observations fell from 32.5% to below 10%.
• Capital deployment rose from 42.9% to 78.0% in the affected test population.

Those are harness numbers. Returns are a separate question, and many agents lost money; the losses are in the paper too. The lesson we took is that an AI trading agent should be judged across the whole path from your instruction to the settled trade. A backtest or a screenshot covers a small part of that path.

What the public benchmarks show

The public tests of frontier models trading real capital with a thin harness mostly lost money, and the write-ups landed on “LLMs can’t trade”. We read the same results differently: the same model in a different harness is a different agent, and the harness is the part those tests did not vary. That is the part DXAP is.

What DXAP does about it

• The model proposes; the policy engine decides. Your loss limits, allowlists and approvals are checked outside the model on every turn. Details on the policy engine page.
• Every turn is on the record, including the turns where the agent decided to do nothing, so you can read why. See debriefs.
• No paper mode. Paper fills at mark with zero slippage and overstates live results. Every agent trades real money, from $10.
• Your wallet. Agents trade through Hyperliquid agent wallets, trade-only keys you can revoke on Hyperliquid at any time.

What this does not mean

None of the above makes trading profitable or safe. Perps are leveraged and agents can lose money. Past performance, including the published research runs, does not guarantee future results. The platform-wide running totals are on the live record, and the datasets behind the research are on dxrg.ai.