Benchmark
Live demo runbook
Reproduce the two-app money-and-message demo on your own simulator.
The hero demo
One natural-language instruction, two apps, about 90 seconds, uncut: the agent sends money in a banking app (with a memo), then reports it in a messaging app — while the permission layer visibly gates the irreversible send until the goal authorizes it.
- 1
Stage a clean simulator
Boot an iPhone simulator with the demo apps installed and reset their state so the run starts fresh.
open -a Simulator - 2
Start recording
Record your screen with the terminal and the simulator side by side (QuickTime → New Screen Recording, or ⌘⇧5). Uncut runs are the proof.
- 3
Run the goal
Hand the instruction to any agent connected to Phone Use — Claude Code shown here.
claude -p "Send $50 to Maya Patel on MyBank with the memo 'dinner', then message Maya on QuickChat that the money is on the way. Confirm both are visible on screen." \ --allowedTools "Bash(bun:*)" - 4
Show the replay
Run the same goal again: known routes replay from the app map with far fewer model calls. Same result, a fraction of the time — that is the memory story in one take.