Make the expectation explicit
Give it unambiguous instructions and inspect the result. Similar labels, hidden state, loading indicators, and unusual layouts can all change what the agent observes.
Put it into practice
Use AuraCheck reports to understand the actual attempt. No single demonstration establishes universal model accuracy, so evaluate with representative flows from your product.
Try this on your next release
Choose one journey, define the expected result, and inspect the evidence from a completed run. Expand the suite after the first test is useful.