Scheduled versus triggered revisits in an agentic trading loop

Revisiting a position sooner can help an agent respond to a changing market. It can also add noise, cost and unnecessary trades. DXAP describes scheduled turns and conditions that wake agents between them. We describe a PROPOSED controlled comparison of those revisit policies. The central question is which timing rule improves the user's specified objective under comparable information and exposure, rather than which arm produces more activity.

Freeze the decision policy and vary the clock

The reference arm runs on a fixed schedule. The treatment uses the same schedule plus a declared trigger condition, such as a price crossing based on a saved reference. Model, mandate, position limits and execution assumptions remain identical. Trigger definitions must be selected from prior data and frozen before the test window. Adjusting the threshold after watching market outcomes would convert the comparison into retrospective selection.

We would log trigger eligibility continuously even when an arm receives no extra invocation. That shared eligibility record establishes which events existed and when. It also lets us compare a triggered revisit with an equally timed observation in a control fixture. Otherwise the treatment receives fresher information and more model calls, making it impossible to tell whether the particular trigger matters.

Count repeated wakeups as a design outcome

An illustrative asset crosses a threshold at 10:01, falls below it at 10:02 and crosses again at 10:03. A naive trigger creates two turns around essentially the same event. The protocol needs an explicit reset condition or debounce rule, together with tests for a trigger firing while a scheduled turn is already processing. We would count merged, suppressed and delayed turns, since those choices affect the actual timing policy.

The outcome record includes observation-only turns. More frequent revisits may improve risk awareness without causing an action, or may induce turnover that harms results after costs. We would measure response delay, missed conditions, decision consistency and trading costs beside portfolio outcomes. Comparing only completed trades would erase an important part of the agent's behavior.

Test timing under stressful periods

The holdout should include quiet windows and rapidly changing markets. Day-level comparisons preserve the fact that many trigger events can share one market shock. We would publish the full threshold settings and their selection history, plus a simple budget-matched control that allocates the same number of extra turns independently of the trigger signal. That control asks whether conditional timing beats merely spending more invocations.

Our continuous record documents historical schedules and agent-created triggers, while the controls paper discusses cadence drift as a harness failure mode. Those findings make revisit timing a substantive frontier question. The proposed experiment would inform how an evolving DXAP harness uses attention: whether a condition delivers a timely, useful reconsideration under the mandate, and whether that advantage survives the additional costs and coordination work it creates.