Evaluate agent decisions together with Android lifecycle behavior.
1. Bounded work
Limit tool calls, concurrent runs and elapsed time. Cancel work when its owner closes unless the product explicitly supports durable continuation.
2. UI instrumentation
Test confirmation dialogs, loading states and restored screens with fake events. Combine these with unit tests for tool validation and a small live-model evaluation set.
3. Device evidence
Measure latency, memory and network use on representative hardware. Model behavior tests alone cannot reveal microphone leaks, lifecycle races or rendering stalls.
Worked scenario
The agent answers correctly in a console test but continues running after leaving the screen. The mobile evaluation must catch that resource leak.
Apply it
Run a matrix of normal, adversarial, cancelled and recreated sessions with observable tool counts.
Check your understanding
You can demonstrate both correct outcomes and bounded mobile resource use. Explain the decision and show evidence from your implementation or design. If you cannot demonstrate it yet, revisit the relevant section before continuing.