Risk limits versus strategy prompts in AI trading
A strategy prompt tells the model how you want it to reason. A policy setting defines a boundary that execution must respect. DXAP separates these roles: the model proposes actions and the policy engine checks them outside the model.
Published by DXAP, a product of DX Research Group. Updated September 13, 2026.
Put preferences in the strategy
Explain the thesis, the evidence you consider meaningful and reasons to stay out. Natural-language instructions are useful for directing judgment and describing exceptions. They are not a reliable substitute for configured execution permissions.
Put enforceable boundaries in settings
Review the available loss limits, asset allowlists and approval rules in DXAP. The policy engine checks proposed actions against configured constraints before execution. Understand what each limit measures and avoid assuming that a policy setting guarantees a particular fill price or caps every possible market loss.
Approve changes intentionally
When adjusting the strategy, distinguish a text clarification from a proposed settings update. Review the actual proposal before approving it. Debriefs can explain why a limit mattered, but an explanation should not silently increase the agent’s authority.
Put it into practice
Instruction: “Prefer a clear explanation over forcing a trade.” Setting: an explicit asset allowlist or approval requirement. Review both, because they solve different problems.
Questions about this workflow
Can the agent talk its way around the policy engine?
DXAP checks proposed actions in a separate deterministic policy engine. The model’s explanation is not itself permission to execute.
Do risk limits eliminate trading losses?
No. Limits constrain permitted actions; market movement, execution conditions and existing exposure can still produce losses.
Read the product details
Run your strategy with DXAP
DXAP is a hosted AI trading agent platform for Hyperliquid traders who want to describe a strategy in plain language, set its limits and inspect its decisions.
No subscription or separate model bill. DXAP charges 2.5 bps (0.025%) on executed agent volume; Hyperliquid fees and funding are separate. Alpha access requires a referral code. The minimum is $10 USDC per agent, with no user paper mode.
Trading can lose money. Examples describe workflows, not expected returns.