Product design question
How would you design Project Mariner (web-browsing agent) to safely book things on a user's behalf?
- Google DeepMind
- 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
Product design for a high-stakes autonomous agent taking real-world actions (payments, bookings) on a user's behalf.
How to approach it
- Clarify the scope: a web-browsing agent completing tasks like booking a flight or purchasing an item using the user's own accounts and payment methods.
- Design explicit consent checkpoints: the agent must show a clear summary of the exact action (price, dates, cancellation terms) and get explicit confirmation before any payment or irreversible booking action.
- Set autonomy boundaries by action type: low-risk browsing and comparison steps can run freely, while payment or irreversible actions always require a final human approval.
- Add transparency during execution: a live, human-readable log of what the agent is doing, so the user can intervene if it goes off track.
- Add safety rails against manipulation: since the agent interacts with arbitrary websites, it must resist deceptive page content designed to trick it into unwanted actions.
- Define success as task completion rate with zero unauthorized charges or bookings, tracked as the primary trust metric above raw speed or convenience.
What a strong answer includes
- Draws a clear, non-negotiable line at payment and irreversible actions requiring explicit human confirmation.
- Names a specific and realistic risk unique to this product: adversarial or deceptive web content trying to manipulate the agent.
- Balances autonomy and safety by tiering permissions by action risk rather than a single blanket confirmation step.
- Prioritizes zero unauthorized-action incidents as the top metric, above speed or task volume, appropriate for a high-stakes agent.
- Adds a live transparency log so users can catch problems mid-task, not just before or after.
Common mistakes
- Allowing the agent to complete payments or bookings without an explicit, final human confirmation step.
- Ignoring the risk of adversarial web content manipulating the agent's actions.
Likely follow-up questions
- How would you handle a website designed to trick the agent?
- What would you do if a user complained about an unauthorized charge?
- How would you balance safety checkpoints against task completion speed?
More product design questions
- Design an enterprise agent marketplace on the Gemini Enterprise platform.Google DeepMind · Product design · Hard
- How would you improve Jules (coding agent) to compete with Cursor and Claude Code?Google DeepMind · Product design · Medium
- Design a consumer feature using Project Astra (real-time multimodal assistant).Google DeepMind · Product design · Hard
- How would you improve Gemini's integration across Google Workspace (Docs, Gmail, Sheets)?Google DeepMind · Product design · Medium
- Design a TV for a car.Google · Product design · Hard
- Design a refrigerator for the blind.Google · Product design · Hard
More questions from Google DeepMind
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