1:1 mentoring with Big Tech AI engineers
Anthropic
S05
HardPremium

Design a Computer-Use / Browser Agent

Design an agent that operates a real browser for a user, filling forms, navigating and booking, while every page it reads may be trying to hijack it.

Computer UseGUI GroundingVisionSafetyRecoveryInjection Defense

Key Requirements

  • Operate a disposable browser behind a domain allowlist
  • Bound what a prompt-injected model can do, in code
  • Hold purchases, sends, account changes and consent for a person
  • Keep logins out of the model's context
  • Verify each step and recover when navigation goes wrong

Interviewer Follow-ups

  • Q1How do you prevent a malicious webpage from hijacking the agent?
  • Q2How do you handle irreversible actions like purchases?
  • Q3How do you verify the agent completed the task correctly?
  • Q4Browser use or computer use: which toolset, and why?
Loading...