✳ AI can click. Should it?
Give it the keyboard.
Keep the keys.
Pick a real task, set the agent’s access, and draw the line between “do this” and “show me first.” Based on published agent docs and benchmark evidence—not vibes.
Run the delegation test
Scenario 01 / 04
What would you hand off?
Agent’s permission
Least privilege: can browse and compare; cannot change anything.
DELEGATE THE RESEARCHLOW SIDE EFFECT
Let the agent do the legwork.
Keep with you
A guardrail, not a guarantee.
What the evidence actually saysCapability is real. So are the limits and the need for control.
369
real web + desktop tasks in OSWorld’s benchmark suite. Task design ↗
Confirm first
Google’s optional safeguards can pause sensitive or irreversible actions. Read the release ↗