Cost-Aware Evaluation for Trading Research Subagents

A research subagent can provide a useful fact and still make the trading turn worse if it arrives too late or consumes resources on an already settled decision. We would evaluate its marginal contribution to the parent agent, rather than grading the worker's prose in isolation.

DXAP publicly describes specialist workers for research and chart analysis within agent turns. That makes a concrete frontier question: which decisions benefit from additional investigation, and when should the parent finish with the evidence already available? The following experiment is proposed, with no claim about measured DXAP subagent returns.

Pair a Turn With Its Additional Investigation

Create two versions of each saved case. The baseline parent receives the ordinary data packet. The treatment receives that packet plus one subagent result produced from the same time-bounded evidence. Keep the parent model, mandate, policy checks, and grading rules fixed.

For an illustrative case, the parent sees a price jump and a social headline. A research worker retrieves a timestamp showing that the headline predates the jump by several hours. The useful contribution is the correction to event timing. A long summary of unrelated news increases input size without resolving the decision's uncertainty.

Another case might require chart analysis to distinguish a temporary wick from a sustained move under the mandate's definition. Evaluate whether the returned finding changes the relevant action or confidence. A correct observation that duplicates the baseline packet has limited marginal value in that comparison.

Count the Cost at the Parent Level

Record worker inference cost, parent input growth, tool requests, and wall-clock delay. Report costs per evaluated turn and per useful correction. A successful worker answer that the parent ignores remains a successful retrieval and an unsuccessful decision contribution; both observations matter.

An illustrative worker costs $0.02 and adds four seconds. Across 100 calls, direct worker cost is $2, before parent processing and tool costs. If ten calls correct a material factual error, the direct cost per correction is $0.20. These arithmetic examples describe accounting, not a return threshold or a measured product result.

Market opportunity costs require their own assumptions. A four-second delay may matter in one task and be negligible in another. Measure the relevant input age and deadline instead of declaring all faster workers superior.

Route by Unresolved Question

Our proposed routing policy invokes a worker when a named uncertainty can affect an admissible action. Compare that policy with always-invoke and never-invoke baselines on a held-out case set. Include failed calls, duplicate findings, and cases where extra information changes a good decision into a worse one.

The harness-transfer method provides the frozen-component comparison. The trace feedback method provides the record linking worker evidence to the parent's proposal and eventual outcome.

A research result would report decision corrections, regressions, coverage, and total cost together. Better worker prose alone supports a narrow quality claim. Better parent decisions under the declared timing and accounting support a stronger contribution claim. That distinction lets the harness evolve toward useful specialist work while keeping its tradeoffs visible.