--- title: "Questions with a shelf life" author: Dan Birman url: https://dbirman.com/shelf updated: 2026-11-09 summary: "An agent that asks a question two hours into a task has already done two hours of work on a guess." --- # Questions with a shelf life Agent Trust Patterns · 03 · v1 · Oct 2026 · 4 min An agent that asks a question two hours into a task has already done two hours of work on a guess. Questions lose value as work goes on, and some lose it fast. Ask early, say what you’ll assume if nobody answers, and once an answer can’t help this run, offer a new run instead. ## The problem Long-running agents either interrupt constantly or never ask at all. Early reviewers of browser agents named the constant pauses as the biggest problem. A 2026 study measured when questions help: questions about the goal lose nearly all their value after about 10% of the work, questions about inputs last until about halfway, and asking after that point does worse than never asking. No frontier model it tested asked at the right time. ## The pattern 1. **Ask everything you can up front.** Before starting, the agent lists the decisions it expects to need, in one place. 2. **Say what you’ll assume.** Every question shows its default, so silence becomes a decision you can see. 3. **Show when each question expires.** Expiry follows progress, not the clock. A question about the goal expires early. A question about a detail lasts longer. 4. **Late answers start a branch, not a U-turn.** After a question expires, answering offers a rerun from the last checkpoint, with the change, instead of changing course mid-task without telling you. ## Try it An assistant is planning a team dinner for eight. It asks four questions up front, then starts. Answer some, ignore others, and try answering one after it expires. [Interactive demo. Switch to the human view to try it.] Made-up task. The run is compressed to about 30 seconds. ## Why I care At Ruli, we interviewed lawyers and watched session recordings. Lawyers who opened the AI chat often gave up quickly. They knew what they needed but not how to phrase it, so the first answer missed, and every follow-up question came too late to save the session. We tried a step-by-step wizard, then the same wizard in its own tab, then a blank box. What worked was asking the structured questions at the very start, at the moment they mattered, and staying out of the way after that. ## The evidence Timing results: [Gulati et al., “Ask Early, Ask Late, Ask Right”](https://arxiv.org/abs/2605.07937). Underspecification comes in kinds, and not every gap matters to the outcome ([Pu et al.](https://arxiv.org/abs/2602.10525)). Anthropic found that on complex tasks, the agent stops itself to ask more often than people interrupt it ([Anthropic](https://www.anthropic.com/research/measuring-agent-autonomy)). The Magentic-UI team notes there is still “no ground-truth signal for when to interrupt users” ([Mozannar et al.](https://arxiv.org/abs/2507.22358)). ## Closest prior work Devin asks clarifying questions up front when it’s unsure ([Cognition](https://cognition.com/blog/devin-2-1)). Claude Code can rewind to checkpoints ([Anthropic](https://code.claude.com/docs/en/checkpointing)). Ambient-agent designs sort interruptions into notify, question, and review. None of them show the assumed default and an expiry on each question, or treat a late answer as a fork. ## Watch for Don’t hard-code the expiry points. The 10% and 50% figures come from simulated tasks and vary by model, so treat them as a starting point to tune per task. And don’t turn defaults into a trap: a default that spends money should still wait for an answer.