Apple Researchers Built a Virtual User to Test AI Assistants Before Real People Do
A new research framework simulates the back-and-forth of real app use, so proactive AI assistants can be tested and scored without putting actual users at risk.
Key points
- Apple ML Research published PARE, a framework for testing AI assistants that act without being asked, in 2025.
- PARE models apps as finite state machines, a structured map of every screen and action an app can offer, instead of flat command lists.
- Current AI assistant testing methods miss the step-by-step, stateful nature of real app use.
- The framework makes it possible to simulate a realistic user so developers can measure how well a proactive agent performs before it touches a live product.
Most AI assistants wait for you to ask them something. A proactive agent, software that watches what you're doing and acts on your behalf before you type a request, is a different animal. It might notice your flight is delayed and rebook it, or spot a bill due tomorrow and schedule the payment. Useful, yes. Also risky if it gets things wrong.
Building and testing these assistants is hard. You need a realistic stand-in for a human user, something that clicks through menus, fills in forms, changes its mind mid-task, the way real people do. Until now, no good tool existed for that.
Apple ML Research introduced PARE, the Proactive Agent Research Environment, to fill that gap.
What does this actually change for developers?
It gives them a safe sandbox to break things before shipping. PARE works by mapping each app as a finite state machine, think of it as a detailed flowchart showing every screen, every path a user can take. That's a much richer picture than the flat command lists that earlier testing tools relied on.
With that map in place, PARE runs a simulated user through realistic sequences: open the calendar, check Tuesday, see a conflict, try to reschedule. The AI assistant being tested has to respond to each step, just as it would with a real person at the keyboard.
Previous frameworks treated apps like a vending machine: put in a command, get an output. PARE treats them like a conversation that unfolds over time, which is far closer to how people actually use software.
What does this mean for ordinary people?
Nothing changes in your apps today. This is research infrastructure, the behind-the-scenes scaffolding that AI teams use to check their work.
The downstream effect matters, though. Proactive agents are coming to phones and productivity tools faster than most people realise. Our story from 14 July found that half of enterprises had already shipped an AI agent that broke something for a real customer after passing internal tests, and only 5% fully trusted the testing meant to catch those failures. A framework like PARE is exactly the kind of structured answer that gap demands.
One honest caveat: a simulated user is still a simulation. Real people do unpredictable things, and no lab environment captures all of that. Good test results in PARE are a necessary step, not a guarantee.
The real question isn't whether proactive agents are coming. They are. What matters is whether the teams shipping them have rigorous ways to measure failure before you see it. PARE is a credible attempt at that, and it's more than most current pipelines offer.
