AI & Technical question

Design guardrails for Manus taking actions on websites that lack APIs.

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 and technical guardrail design for an agent operating in the riskiest environment: websites with no structured API to constrain its actions.

Identify the core risk: without an API, the agent interacts via clicking and form-filling, which can trigger unintended actions like purchases, account changes, or destructive operations if misread.
Classify actions by risk before executing: read-only actions (browsing, searching) proceed freely, while any action matching patterns like 'submit', 'buy', 'delete', or 'confirm' triggers a required checkpoint.
Add a sandboxed dry-run step where possible, simulating the click's likely effect before committing to it, to reduce misclick risk on ambiguous page elements.
Add domain and action allowlists for the highest-risk categories, such as requiring explicit user permission before the agent interacts with a payment or account-settings page at all.
Add a rate limit and anomaly detection layer to catch an agent stuck in a loop or behaving unexpectedly on a page it may be misinterpreting.
Measure success as zero unauthorized high-risk actions in testing and production, tracked as the top guardrail metric above task success rate.

Names the specific structural risk (no API means action intent must be inferred from page content, which is inherently riskier).
Proposes a concrete risk-classification and checkpoint system based on action keywords and page context.
Adds a dry-run or simulation step, a specific technical mitigation for misclick risk on unfamiliar pages.
Includes anomaly detection to catch the agent behaving unexpectedly, not just individual risky actions.
Prioritizes a zero-unauthorized-action guardrail above raw task completion, appropriate for this risk level.

Treating all web actions as equally safe just because there is no formal API involved.
Not addressing how the agent handles ambiguous or deceptive page elements.

Chapters of the AI PM course that teach what this question tests.