Product design question

How would you design Comet Assistant to safely act on tasks like booking flights and managing email?

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

AI agent product design balancing genuine usefulness against the real risk of an autonomous agent taking costly, irreversible actions.

Segment tasks by risk: informational research is low risk, booking a flight or sending an email on the user's behalf carries real financial and reputational consequences if wrong.
Propose a tiered autonomy model: the agent can research and prepare an action, like a drafted email or a flight itinerary, but requires explicit user confirmation before executing high-stakes actions.
Design a clear preview step showing exactly what the agent is about to do, in plain language, before the user approves, avoiding vague or overly technical confirmations.
Address failure handling: if the agent hits an unexpected state mid-task, like a price change during booking, it should pause and re-confirm rather than proceeding on stale assumptions.
Define success: task completion rate for approved actions, and, critically, a near-zero rate of unintended high-stakes actions executed without explicit confirmation.

Proposes a clear tiered permission model distinguishing research from execution, rather than treating all agent actions with the same level of autonomy.
Designs the confirmation step to be genuinely understandable, a plain-language preview, not a buried technical detail users will click through without reading.
Names stale-state handling explicitly, like a changed flight price mid-booking, as a realistic failure mode that needs its own safeguard.
Sets a safety-first success metric, near-zero unintended high-stakes actions, as more important than raw task completion rate.

Designing uniform autonomy for all tasks regardless of risk, ignoring the real difference between research and irreversible financial actions.
Ignoring failure and stale-state handling, assuming the world stays static during a multi-step agentic task.

How would you handle a user who wants full autonomy without confirmation prompts?
What would you do if the agent made a booking error despite user confirmation?