Cognition gave Devin a Mac this morning. That sentence sounds small until you remember what’s been broken about coding agents for two years: they can all write Swift, run the build, and iterate until the compiler goes green. None of them could open the app. Compiling is a fact about your code. Working is a fact about your app, and checking it requires playing with the thing.
Their own example is an iPhone game that should resume where the player paused it. To verify that, an agent has to play, pause, quit, relaunch, and look. So they gave it a full macOS VM on bare-metal AWS hardware, a warmed-up Xcode, an iOS Simulator, and eyes.
I read the whole engineering post so you don’t have to. Four things stood out.
They fought Apple’s networking and won ugly. Apple’s built-in NAT kept overwriting the same macOS firewall that Cognition uses to enforce session access rules, so two systems raced to configure one packet filter. Their fix was to drop down a layer and write their own user-space Ethernet gateway that answers ARP requests by hand. That’s a lot of work to guarantee “the agent can only reach what you allowed,” and it tells you enterprise security reviews are driving the roadmap.
They pre-click the permission dialogs. A fresh macOS install stops to ask about accessibility and screen recording, and every dialog is a spot where the agent sits waiting for a human. Cognition mounts the disk image and edits macOS’s internal configuration databases directly, then boots it once to warm Xcode and the Simulator before you ever arrive. Anyone who’s waited out a cold Xcode first-launch knows how much dead time that removes.
Devin reads the app instead of staring at it. Screenshotting the screen and asking a vision model “where’s the Resume button” is slow and expensive at scale. Devin queries the accessibility tree instead, the same structure built for assistive tech, which hands back every control’s name, role, state, and available actions as text. Screenshots stay for the stuff that’s drawn rather than declared, like gameplay itself.
The quiet acquisition paid off. The accessibility layer is built on tooling from Dioxus, a company Cognition bought with close to zero coverage. Turns out that was the piece making agentic computer-use affordable on Apple platforms.
And the launch note itself is the workflow pitch: Devin builds on its own Mac, sends you a screen recording in Slack, and hands you a TestFlight link. You go to bed with a queue and wake up with a build on your phone.
Now, my skepticism, because Cognition earned some. The 2024 launch demo was later shown to be less autonomous than it looked, and this post is heavy on capability and silent on success rate. There’s no pass rate on iOS tasks anywhere in it, no comparison to a human running the same loop. A Mac with accessibility access makes autonomous iOS work possible, and possible is where the marketing usually stops. My bet: month one shows Devin is genuinely good at verifying behavior and close to useless at judging design, because the accessibility tree describes a layout without ever finding it ugly.
Which means the winners here will be the people who hand it the right kind of work, framed the right way. That framing is a skill, I’ve spent the day building the system for it, and it’s below.
The free section ends here. Everything below is the working system.
Premium subscribers get the Devin Mac Handoff System: eleven prompt templates covering the full iOS verification loop, the accessibility-tree query patterns that cut a screenshot-heavy session by roughly two-thirds, a session setup checklist for your first Mac workspace, the cost model that tells you when a Mac VM beats running Xcode locally, and a decision table comparing Devin against Claude Code, Cursor, and Codex on eight mobile tasks.
THE DEVIN MAC HANDOFF SYSTEM
Part 1: Before you write a single prompt
Six setup moves for your first Mac session. I’d love to tell you I derived these from first principles, but really they’re the six ways my first sessions went sideways.