Skip to content
DXAPDOCS

Research behind DXAP

DXAP is a working product and an ongoing research program. We want to push what agents can actually do, and that means studying the whole system: the information an agent sees, how it reasons, which tools it can use, what gets executed, and what happens afterward.

Early participation matters here. A strategy that behaves strangely, a useful new tool, a failed exit, or a very specific piece of feedback gives us something concrete to investigate. We combine those questions with our market-data infrastructure and recorded decision paths to develop datasets, train and optimize models, and test new specialized agents and approaches.

How the work connects

Section titled “How the work connects”
  1. Participate

    Strategies, real outcomes and feedback from early users.

  2. Observe

    Market data and linked decision, execution and outcome records.

  3. Build

    Datasets, model training, specialized agents and system changes.

  4. Evaluate

    Replays and controlled comparisons, including failures.

  5. Release

    Measured improvements, new tools and the next experiment.

Each release gives participants new things to test and gives us the next questions to study.

Participation gives us questions and evidence. We study the available market and execution records, build model and system candidates, compare them in replays and controlled evaluations, then bring useful improvements back into the product. New releases create the next round of experiments.

The lab trains and evaluates model candidates; your running agent uses the model and configuration you’ve chosen. We study the models and their surrounding system. Changing context, memory, tools, execution or a model can each change behavior; we try to distinguish which part made the difference.

What we’re working toward

Section titled “What we’re working toward”
  • Model training and optimization: stronger decision-making, more useful specialized models and subagents, and better use of the information available at a turn.
  • The operating system around the model: strategy interpretation, memory, market context, tool use, execution and recovery. A stronger model still needs a system that gives it the right problem and can carry out its decisions.
  • Data and evaluation: turning linked records into reusable scenarios, comparing candidates on the same problems, and checking outcomes after fees and execution effects.
  • New agent types and environments: extending the work beyond today’s Hyperliquid perpetuals, with different markets and coordination problems that deserve their own experiments.

These are research directions. New models, markets and capabilities will get their own availability announcements and evidence. A historical study or a promising training run is not a performance report for the current DXAP Alpha.

Read the work

Section titled “Read the work”
Resource What you’ll find
DXRG research library Deployment accounts, methods, evaluation protocols and versioned public datasets.
Operating-Layer Controls for Onchain Language-Model Agents Under Real Capital The DX Terminal Pro research record: how user mandates, prompts, controls and execution fit together in a 21-day real-capital deployment.
What LLM Trading Agents Actually Do in Production A six-month record from two fleets, studying behavior, sizing, exits and model comparisons. Its DXAP observation window ends August 15, 2026, before the current Alpha.
Paper artifacts and aggregate data Figures, methods, datasets and the discussion guide behind that production study.
Research field notes Questions, methods and examples for building and assessing trading agents.

The published record helps explain why we work on the harness and evaluation alongside the models. It also preserves failures and limits instead of treating every interesting result as an improvement. Forward results from the current system will be measured separately.

Help us find the next question

Section titled “Help us find the next question”

Share what your agent was trying to do, what it actually did, and the trade or Activity entry that shows it. Comparisons across strategies, models and market conditions are especially useful. You don’t need to turn every observation into a theory; sometimes “this feels wrong” is the right starting point.

Season 0 is how we’re organizing this early participation. Bring the interesting cases to Discord, use the app’s feedback button, and keep sharing your results.