Product design question

Design an AI agent that can take actions on behalf of users. How would you define its permissioning and control model?

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

Ability to design a permission system for autonomous agents, balancing usefulness against the risk of unwanted or harmful actions.

Clarify what actions the agent takes, since risk and permissioning differ between reading data, sending messages, and purchasing.
Classify actions by reversibility and blast radius, for example reading data versus sending money.
Define a tiered model: read only by default, explicit approval for irreversible or costly actions, autonomy only for low risk repeatable tasks.
Design confirmation UX that previews what the agent will do, not just an allow prompt.
Add hard limits such as spend caps, rate limits, an audit log, and a kill switch to halt an agent mid task.
Decide how permissions are scoped and revoked, for example per integration or per session with expiry.

Separates permissioning by reversibility rather than one blanket toggle.
Proposes concrete defaults: read access on, write access opt in, spend capped until the user raises it.
Includes an audit trail and undo path as a control, not just a pre action gate.
Flags that over prompting for confirmation defeats the point of an agent, so batches low risk approvals.

Proposing a single all or nothing permission toggle instead of tiers by risk.
Skipping the audit and undo mechanism, which matters as much as the upfront gate.

How would you handle an agent that chains several actions across different services?
What happens when the agent is uncertain whether an action needs approval?