Proactive Agent Research Environment: Simulating Active Users to Evaluate Proactive Assistants
Apple
Apple-backed researchers built a fake app world to test AI assistants that act before you ask. It's called Pare, and it finally treats apps like apps instead of glorified API lists.
Based on reporting by Apple — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Every AI assistant demo promises the same thing: it'll notice what you need and just do it, no prompting required. The problem is nobody has had a good way to test whether that actually works, because most simulation benchmarks treat an app like a bag of API calls you can fire off in any order. Real apps don't behave that way. You have to open a screen, wait for a state to load, navigate a menu, then take an action — and the options available to you change depending on where you are. A team from UC Santa Barbara, working with researchers at Apple and the University of Washington, decided that gap was worth fixing.
Their answer is Pare, short for Proactive Agent Research Environment. Instead of modeling apps as flat toolkits, Pare represents each one as a finite state machine, so a simulated user actually has to navigate screens and states the way a person would, with the available actions shifting as context changes. That stateful design is what lets the framework support something genuinely new: an active user simulator, one that behaves like a person going about their day rather than a static task description waiting to be executed.
On top of that scaffolding sits Pare-Bench, a set of 143 tasks spread across communication, productivity, scheduling, and lifestyle apps. The tasks aren't simple single-step requests. They're built to probe four specific skills a proactive agent needs: reading context correctly, inferring what the user actually wants without being told outright, knowing when to step in versus when to hold back, and juggling multiple apps to pull off a task that spans more than one tool.
That last point matters more than it might sound. Plenty of agent benchmarks still measure how well a model calls a single API correctly. Pare-Bench is explicitly testing timing and restraint, arguably the harder and less glamorous problem — an assistant that jumps in too early or too often is arguably worse than one that does nothing. The paper notes that even sophisticated benchmarks like AppWorld see agents fail more than half their tasks, which gives a sense of how far current systems are from reliably operating inside messy, sequential, real-world app states.
My take — AI-written commentary, not fact-checked reporting
I've been saying for a while that the 'proactive assistant' pitch has been running way ahead of any credible way to measure it, so it's refreshing to see someone build the boring infrastructure instead of another demo video. The real tell here is that Apple let this ship as an academic-style release rather than a product teaser — which tells you they know proactive agents are still years, not months, from shipping reliably on your phone.
Read more about this at: Apple