Agent Clarification Economics: When Asking One Good Question Beats Taking Ten Risky Actions
Lab Bench — July 13, 2026. The best agent is not the one that asks least. It asks when uncertainty is expensive: before a consequential choice, when a default is likely wrong, or when a short question avoids a much larger cost of error.
Lab Premise
Clarification is a decision under uncertainty. Asking too often adds user effort and delays useful work; asking too little can turn a reversible draft into an irreversible mistake. A capable agent should identify the smallest question that materially changes the safe action path, then proceed with transparent defaults when the decision is low-risk and reversible.
Failure Surface
| Condition | What Goes Wrong | Reliable Behavior |
|---|---|---|
| Ambiguous target or audience | Agent commits work to the wrong person, market, or system. | Ask one scope-defining question before consequential execution. |
| Low-risk reversible detail | Conversation stalls over a preference that can be changed later. | Choose and disclose a sensible default, then proceed. |
| Urgent task with incomplete data | Delay costs more than a bounded assumption. | State the assumption, limit the action, and preserve an easy correction path. |
| High cost of error | Wrong assumption triggers loss, exposure, or irreversible change. | Pause for a specific confirmation rather than guessing. |
Benchmark Design
Build a 20-case clarification lab with cost-of-delay and cost-of-error scoring. Vary ambiguity, reversibility, urgency, user effort, and the value of an available default. Include tasks where a question is essential, tasks where one crisp question is enough, and tasks where asking is counterproductive.
- Record the uncertainty the agent identifies and its estimated consequence if wrong.
- Measure the specificity and decision value of every question.
- Score whether the agent proceeds safely after an answer or justified default.
Scorecard
| Metric | Evidence of Good Judgment |
|---|---|
| Ambiguity detection | Flags uncertainties that actually alter the action path. |
| Question value | Each question is narrow, answerable, and materially decision-relevant. |
| User-effort control | Avoids preference surveys and unnecessary multi-question blocking. |
| Default quality | Uses reversible, disclosed assumptions for low-cost uncertainty. |
| Action threshold | Escalates before high-cost or irreversible error. |
Practical Verdict
One good question can beat ten risky actions. The reliable agent prices uncertainty correctly: it asks before expensive mistakes, defaults when the cost is low, and never disguises a consequential guess as certainty.
Methodology Note
These lab frameworks assess observable behavior under controlled scenarios. Passing requires a traceable decision process, appropriately bounded uncertainty, and repeatable evidence—not merely a plausible final answer.