OpenAI · AI product · Mid · 45 to 60 minutes

How should OpenAI design a deep research agent?

OpenAI builds AI models and products that help people reason, create, and complete complex tasks. Answer this product management case and get it graded line by line against a weighted rubric.

Product strategy
AI systems thinking
Prioritization
Experiment design
Risk management

The situation

The product lead for applied AI at OpenAI has asked you for a recommendation on a new deep research capability. The team wants to help users answer questions that cannot be solved in one shot, such as investigating a topic across many sources, forming a plan, and producing a trustworthy synthesis. Interest is high because users increasingly want AI that can do more than provide a quick response, but there is real concern about launching something that appears impressive while failing quietly on hard tasks. The debate is not only about what tools the agent should have, but how it should break work into steps, decide what to verify, and react when evidence is weak, conflicting, or hard to check. Verification can be expensive and, for some claims, nearly as difficult as the original research itself. You need to recommend the product approach for an initial release, including how the system should plan, verify intermediate findings, and decide when to continue, revise its plan, ask for clarification, or stop.

What to produce

Write a 1 to 2 page product strategy memo.

Your answer must

Define the primary user and job to be done, and clearly distinguish deep research from simple question answering.
Propose an end to end agent workflow that covers planning, information gathering, synthesis, verification, and when outputs from one step should change the next step.
Explain how the product should handle verification, including what to verify, how much effort to spend, and what the system should do when verification is inconclusive or too hard.
Recommend the MVP scope and the main product trade offs, including accuracy, latency, cost, and user trust.
Specify 3 to 5 success metrics or evaluation methods for the initial release, with at least one focused on reliability or trustworthiness.
Identify the top failure modes and propose mitigations, including how the user experience should communicate uncertainty or limitations.

Constraints

Assume the agent can use OpenAI models, browsing, and retrieval over user provided documents, but no routine human review in the loop.
The system must work despite model mistakes, incomplete sources, and occasional tool failures.
Verification effort is limited, so the product cannot fully audit every claim on every task.
The first release should be useful across a broad set of knowledge work tasks, not only one narrow industry.

Format

Use sections for user and problem, proposed workflow, verification approach, MVP scope and trade offs, metrics, and risks.

How OpenAI actually played it

A worked AI PM interview answer on designing a deep research agent from first principles, covering the plan and verification steps and how each step's result feeds back into the plan.

Use first-principles thinking on agent design
Loop verification back into planning
Treat verification difficulty as a design constraint

Built from AI PM Interview Question | Design a Deep Research Agent by Technomanagers.