Product design question
How would you design Comet Assistant to safely act on tasks like booking flights and managing email?
- Perplexity
- Product design
- Hard
Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.
Start a mock interview on this question · Mock interview from a job description
What this question tests
AI agent product design balancing genuine usefulness against the real risk of an autonomous agent taking costly, irreversible actions.
How to approach it
- 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.
What a strong answer includes
- 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.
Common mistakes
- 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.
Likely follow-up questions
- 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?
More product design questions
- How would you improve Perplexity's Comet browser for daily use?Perplexity · Product design · Medium
- Design an AI shopping assistant inside Comet.Perplexity · Product design · Medium
- Design a 'Deep Research' feature that produces trustworthy, citable reports.Perplexity · Product design · Hard
- Design the core interaction model for a Project where a human and multiple agents collaborate on ongoing work. How would you handle delegation, approvals, edit history, permissions, and recovery from bad agent actions so the experience feels powerful but dependable?Perplexity · Product design · Hard
- Users of Perplexity Computer say they do not understand why the agent clicked, searched, or changed course, and they hesitate to let it continue. Design a core interaction model that makes the agent’s intent, current state, and next actions understandable and controllable without turning the product into a complex dashboard.Perplexity · Product design · Hard
- Perplexity notices that many users start a research task on web, ask several follow-ups, but do not return to that work on another platform later. How would you identify the single highest-leverage friction point in the begin → continue → return journey, and decide whether to invest first in context carryover, navigation, or collaboration?Perplexity · Product design · Hard
More questions from Perplexity
Learn the skill behind it
Chapters of the AI PM course that teach what this question tests.
- Chapter 4: Discovery and strategy for AI products
- Chapter 7: AI UX and human oversight: design for a system that is wrong sometimes
- Chapter 14: Get the job: the AI PM interview loop