Cheap analysis makes owner intent more consequential

When routine analysis becomes cheap, the owner has more consequential work to do. Someone must decide which question matters, what outcome the agent should pursue and which costs are acceptable. Our 2026–2031 scenario is that judgment shifts toward defining intent and revising mandates as circumstances change. More analysis increases the number of plausible actions, making the selection problem harder to ignore.

A capable agent can explain why several strategies appear reasonable. It cannot derive an owner's preferences from market prices alone. An owner who values a stable exposure has a different problem from one seeking a speculative short-term opportunity. Even when both receive the same probability estimate, the appropriate action can differ because the purpose of the account differs.

This shift is especially visible in midhorizon decisions, meaning forecasts and holding choices over hours to days. Routine work can gather public information and compare scenarios within that window. The owner still decides whether the resulting opportunity belongs in the mandate. A technically sound analysis can answer a question the owner had no reason to ask.

The ambiguous request

Consider an illustrative request: “Find good opportunities while keeping risk low.” The words leave several conflicts unresolved. Is the priority avoiding large exposures, reducing turnover or preserving cash for an upcoming obligation? Would the owner accept a longer holding period to avoid trading costs? Does a seemingly attractive opportunity deserve capital if it increases an exposure already held elsewhere?

A useful agent makes these conflicts concrete. It can offer alternative interpretations and show how each changes permissible behavior. The owner supplies the preference. An explanation becomes actionable only after that preference has a clear place in the mandate and, where supported, in the configured controls.

The DXAP creation guide asks owners to specify markets, action conditions and reasons to wait. It also distinguishes writing a limit in strategy text from setting an actual policy field. That distinction gives intent a practical form: prose guides interpretation, while configured policies constrain particular execution requests.

The chat and instructions guide describes a proposal and confirmation flow for settings and persistent instructions. Current chat can support a discussion of intent, with approval remaining a separate action. Our projection is that interfaces of this kind become more important as agents produce more competent alternatives, rather than that any particular future feature is scheduled.

Evaluate preference preservation

We would test the thesis by holding the evidence constant and changing the owner's stated objective. Use a saved market and account snapshot, then construct several explicit mandates: preserve an existing exposure, pursue a bounded directional opportunity, or wait until a specified uncertainty resolves. Freeze model access and available tools. Each mandate should include a clear priority when two instructions conflict.

The evaluation asks whether proposed actions change in the direction implied by the owner's objective. A decision to reduce exposure may fit one mandate while violating another. Measure unsupported preference invention, questions that expose genuine ambiguity and compliance with configured restrictions. Assess forecast quality separately, using the same target and information cutoff across mandates.

This design avoids a common mistake: ranking every owner by one aggregate return figure. A mandate can reasonably sacrifice expected return for a different objective. Evaluation should reveal that tradeoff rather than silently substitute the evaluator's preference. Simulated economic outcomes remain relevant, but they describe the cost of a chosen objective as well as the quality of the decision process.

The proposed comparison also needs an ambiguous-request arm. If an agent acts confidently under contradictory instructions, greater analytical competence has failed to preserve the owner's authority. If it clarifies every trivial detail, the service imposes too much supervision. The useful boundary is the set of unresolved preferences capable of changing a material action.

Our scenario would weaken if owners consistently prefer fixed standardized objectives and rarely benefit from choosing among alternatives. It would also weaken if agents can reliably infer those preferences from authorized context with little confirmation. Either finding would narrow the role of explicit mandate design. We would still need to know how the owner corrects an inference when circumstances change.

The judgment that remains scarce is specific: deciding what the account exists to do, recognizing when that purpose has changed and accepting the resulting tradeoffs. Better research helps by making those choices legible. It can turn a vague preference into alternatives whose consequences the owner can inspect, giving human intent a stronger role in the decision.