Cognition launched macOS support for its agent Devin. It works in virtual Mac machines in the cloud and can launch the app, close it, open it again and check how it behaves.
- The virtual Macs run on AWS EC2 Mac with Apple's Virtualization.framework.
- The images come with Xcode ready, a warmed-up simulator and permissions granted, so the agent does not wait for a human.
- For buttons, Devin reads the app's accessibility tree, not only screenshots.
It compiles. The most deceptive sentence in mobile development.
An app can compile cleanly and still not remember where the player stopped when you close and reopen it. Checking that takes hands: you play, pause, quit, launch again and look. According to Cognition, most agents so far stopped at the compiler.
The best part of their write-up is not marketing. They turned down Apple's managed network setup because it fights their own firewall rules - two systems touching the same thing. Instead of hoping one would win, they wrote their own gateway. More work, but right.
And the accessibility tree. The layer built for people with impaired vision turns out to be a good way for a machine to understand an interface without guessing from pixels. Accessibility done well has long paid bills nobody expected.
If you build apps for iPhone and have tested them by hand, you now have an agent you can tell “check that it remembers the pause”. Whether it does that reliably on your project you will only find out by giving it your project.