Task

Try once

“Can my user actually do this?”

Ad-hoc. You’re building a new flow and you want to know if it works for a real user, without scheduling a research session. Write the task, watch the agent try, read the screenshot of where it got stuck.

Experiment

Compare two builds

“Which design wins?”

Run the same task against two builds of your app, side by side. See which one your users can complete and which one breaks, before the design debate goes to A/B testing in production.